Foundations of Software Science and Computation Structures, 19 conf., FOSSACS 2016 9783662496299, 9783662496305


250 18 7MB

English Pages 548 Year 2016

Report DMCA / Copyright

DOWNLOAD PDF FILE

Table of contents :
ETAPS Foreword......Page 6
Preface......Page 8
Organization......Page 9
Contents......Page 10
Types......Page 13
1 Introduction......Page 14
2 Comprehensive 2 Fibrations......Page 16
3 Comprehensive 2 Parametricity Graphs......Page 19
4 A Type System for Relational Reasoning......Page 21
5 Direct-Image and Pseudograph Relations......Page 25
6 Consequences of Parametricity......Page 26
7 Related and Further Work......Page 28
References......Page 29
1 Introduction......Page 31
2 Guarded Dependent Type Theory......Page 33
2.1 Fixed Points and Guarded Recursive Types......Page 35
2.2 Identity Types......Page 36
3 Examples......Page 37
4 Coinductive Types......Page 39
5 Example Programs with Coinductive Types......Page 41
6 Soundness......Page 43
7 Related Work......Page 44
References......Page 45
1 Introduction......Page 47
2 A Dependently-typed Effectful Language......Page 49
3.1 Fibred Category Theory Preliminaries......Page 52
3.2 Interpretation of Our Language in Fibred Adjunction Models......Page 54
3.3 Fibred Adjunction Models Based on the Families Fibration......Page 57
4.1 Algebraic Effects in the Syntax......Page 58
4.2 Algebraic Effects in the Semantics......Page 59
5.2 Domain-Theoretic Semantics for Recursion......Page 60
6 Conclusions and Future Work......Page 62
References......Page 63
1 Introduction......Page 66
2 Second Order Game Semantics......Page 67
2.1 Examples......Page 68
2.2 Legal Sequences and Strategies......Page 69
2.3 Instantiation......Page 70
3 Generic Strategies as Dinatural Transformations......Page 71
4 Semantics of Bounded Quantification......Page 72
5 A Stateful Language with Bounded Quantification......Page 74
5.1 Denotational Semantics......Page 76
5.2 Semantics of Terms......Page 77
6 Full Abstraction......Page 79
7 Further Directions......Page 80
References......Page 81
Recursion and Fixed-Points......Page 82
1 Introduction......Page 83
2 Background......Page 84
2.1 Joins and Compatibility......Page 87
2.2 Categories of Partial Maps......Page 88
3 As CPO-categories......Page 89
3.1 Reversible Fixed Points of Morphism Schemes......Page 90
3.2 Algebraic -compactness for Free!......Page 92
4 As Unique Decomposition Categories......Page 94
5 Conclusion......Page 97
References......Page 98
1 Introduction......Page 101
2 Bar Recursion......Page 103
3 Continuous Functions on Streams......Page 105
3.1 Individual Tabulations......Page 108
3.2 Global Tabulation......Page 109
4 Bar Induction......Page 111
References......Page 115
1 Introduction......Page 117
2 Preliminaries and Notation......Page 119
3 The Locally Finite Fixpoint......Page 120
4 Instances of the Locally Finite Fixpoint......Page 124
4.1 Generalized Powerset Construction......Page 125
4.2 The Languages of Non-deterministic Automata......Page 126
4.3 The Behaviour of Stack Machines......Page 127
4.4 Context-Free Languages and Constructively S-Algebraic Power Series......Page 128
4.5 Courcelle's Algebraic Trees......Page 130
5 Conclusions and Future Work......Page 132
References......Page 133
1 Introduction......Page 136
2 Notation and Elementary Concepts......Page 138
3 The Intuitionistic Propositional -Calculus......Page 139
4 Strong Monotone Functions and Fixed-Points......Page 141
5 A Digression on Fixpoints and Bisimulation Quantifiers......Page 143
6 The Elimination Procedure......Page 144
6.1 Strongly positive Elimination......Page 145
6.2 Weakly Negative Elimination......Page 147
7 Upper Bounds on Closure Ordinals......Page 148
References......Page 150
Verification and Program Analysis......Page 152
1 Introduction......Page 153
2 The Language......Page 155
3 Monitor Instrumentation......Page 156
4 Monitor Preorders......Page 159
5 Characterisation......Page 163
6 Conclusion......Page 166
References......Page 167
1 Introduction......Page 170
2 Idealized Algol......Page 172
3 Games......Page 173
4 Upper Bounds......Page 176
5 Lower Bounds......Page 182
6 Conclusion......Page 185
References......Page 186
1 Introduction......Page 188
2 Motivating Examples......Page 189
3 Control Flow Graphs......Page 191
4.1 Stores and Distributions......Page 193
4.3 One-Step Reduction......Page 194
4.4 Multi-step Reduction and Loops......Page 195
5 Conditions for Slicing......Page 196
7 Computing the (Least) Slice......Page 199
8 Extensions and Future Work......Page 202
9 Conclusion and Related Work......Page 203
References......Page 204
1 Introduction......Page 205
2 Parameterized Communicating Automata......Page 208
3 Split-Width......Page 211
4 Classes of Bounded Split-Width......Page 214
5 Tree Interpretation......Page 216
6 Further Results......Page 219
References......Page 220
Automata, Logic, Games......Page 222
1 Introduction......Page 223
2.1 Weighted Concurrent Games......Page 226
2.2 Equilibria Notions......Page 228
3 Stationary Strategies......Page 229
4.1 Deviator Game......Page 231
5.1 Multidimensional Objectives......Page 233
5.2 Correctness of the Objectives for Robustness......Page 234
6 Fixed Coalition Game......Page 235
7 Hardness......Page 237
References......Page 238
Quantifier Alternation for Infinite Words......Page 240
1.1 The Quantifier Alternation Hierarchy of First-Order Logic......Page 242
1.2 Decision Problems......Page 243
2.1 Semigroups and Wilke Algebras......Page 245
2.2 Logical Preorders......Page 247
3 Chains for Omega-Languages......Page 248
3.1 i-Chains and Separation for i......Page 250
4 A Separation Algorithm for Sigma2......Page 251
5 A Membership Algorithm for B-Sigma2......Page 253
6 A Separation Algorithm for 3......Page 254
7 Conclusion......Page 255
References......Page 256
1 Introduction......Page 258
2 Nested Words and Nested Word Automata......Page 261
3 Synchronizing Words for NWA......Page 262
4.1 Binary Tree Representation of Nested Words......Page 265
4.2 From Nested Word Automata to Tree Automata......Page 266
4.3 Pebble Games and Strahler Numbers......Page 267
4.4 Bounded Nonemptiness for Implicitly Presented Tree Automata......Page 269
5.1 Small-Cost Synchronizing Words in DFA......Page 270
5.2 Reduction to Short Synchronizing Nested Word......Page 271
References......Page 273
1 Introduction......Page 275
2 Semantics and Undecidability of LTL"3223379 qo......Page 279
3 Nested Counter Systems......Page 281
4 From LTL"3223379 tqo to NCS and Back......Page 285
5 Conclusion......Page 288
References......Page 289
1 Introduction......Page 291
2.1 Words, Strings and Data Words......Page 293
2.3 Transduction of Data Words......Page 294
2.5 Properties......Page 295
2.6 MSO k-types......Page 296
3 Two-Way Transducers on Data Words......Page 297
4.1 Streaming String Transducers with Data Variables and Parameters......Page 300
4.2 From Two-Way to One-Way Transducers......Page 302
4.3 From One-Way Transducers to MSO......Page 304
References......Page 306
Probabilistic and Timed Systems......Page 307
1 Introduction......Page 308
2.1 Labelled Markov Decision Processes......Page 310
3 Undecidability Results......Page 312
4.1 Pure Memoryless Strategies......Page 315
4.2 Memoryless Strategies......Page 317
5 Bisimulation......Page 319
References......Page 322
1 Introduction......Page 324
2.1 Markov Decision Processes......Page 326
2.2 VASS-MDPs......Page 327
2.3 Verification Problems for VASS-MDPs......Page 328
2.4 Undecidability in the General Case......Page 329
2.5 Model-Checking -calculus on Single-Sided VASS......Page 330
3.1 Undecidability in Presence of Deadlocks......Page 331
3.2 Sure (repeated) Reachability in Deadlock-Free P-VASS-MDPs......Page 332
3.3 Almost-Sure and Limit-Sure Reachability in Deadlock-Free P-VASS-MDPs......Page 333
4.2 Almost-Sure Problems in 1-VASS-MDPs......Page 334
4.3 Limit-Sure Reachability in 1-VASS-MDP......Page 336
5 Conclusion and Future Work......Page 338
References......Page 339
1 Introduction......Page 340
2 A Zoo of Timed Temporal Logics......Page 342
3.1 CTMTL Games......Page 344
4 Satisfiability Checking of Counting Logics......Page 350
5 Discussion and Related Work......Page 355
References......Page 356
1 Introduction......Page 358
2 Distributed Interactive Markov Chains......Page 361
2.1 Schedulers and Strategies......Page 362
2.2 Probability of Plays......Page 363
3 Schedulers Are Not that Powerful......Page 364
4 Undecidability Results......Page 365
4.2 Reduction from DEC-POMDP......Page 366
4.3 Undecidability of Qualitative Existence in DEC-POMDP......Page 367
5 Decidability for Non-urgent Models......Page 369
6 Discussion and Conclusion......Page 372
References......Page 373
Proof Theory and Lambda Calculus......Page 375
1 Introduction......Page 376
2.1 Flows and Wirings......Page 378
2.2 Representation of Words and Programs......Page 381
2.3 The Stack Semiring......Page 382
3 Pushdown Automata and PTIME Completeness......Page 384
4 Nilpotency in Stack and PTIME Soundness......Page 387
5 Unary Logic Programming......Page 388
6 Perspectives......Page 390
References......Page 391
1 Introduction......Page 393
2 Modal Logics and the Nested Sequent Calculus KN......Page 395
3 The Focused Systems KNwF and KNF......Page 397
3.1 Cut Elimination......Page 400
3.2 Completeness......Page 402
4 The Synthetic System......Page 404
4.1 Synthetic Substructures......Page 405
4.2 Cut Elimination......Page 406
5 Perspectives......Page 408
References......Page 409
1 Introduction......Page 411
2.1 Non-deterministic -calculus......Page 413
2.2 Resource Calculus and Taylor Expansion......Page 414
2.3 Finiteness Structures Induced by Antireduction......Page 417
3 Strongly Normalizing Terms Are D+ Typable......Page 419
4 D+ Typable Terms Are Finitary......Page 420
5 Finitary Terms Are Strongly Normalizing......Page 423
References......Page 425
A Appendix......Page 427
2 Call-by-name and Call-by-need......Page 429
3 Non-idempotent Intersection Types......Page 432
4 Characterization of name-normalizing Terms......Page 434
5 Characterization of need-normalizing Terms......Page 435
6 Soundness and Completeness......Page 439
7 Conclusion......Page 440
References......Page 442
Algorithms for Infinite Systems......Page 445
1 Introduction......Page 446
2 Model......Page 448
3 Simple Ideals......Page 452
4.1 Accelerations......Page 455
4.2 Coverability Trees......Page 456
5 Complexity Bounds......Page 457
5.1 Fast-Growing Complexity......Page 458
5.2 Upper Bounds......Page 459
6 Concluding Remarks......Page 460
References......Page 461
1 Introduction......Page 463
2 Summary......Page 465
3 Challenges and Techniques......Page 467
4 Preliminaries......Page 468
5.1 Proof Overview and Notation......Page 470
5.2 Normal Paths......Page 471
5.3 Length of Shortest Paths......Page 475
References......Page 478
1 Introduction......Page 480
1.1 Attacking the Invariance Problem......Page 481
2.1 Setting......Page 483
3.1 Finiteness of the Set of Essential Words......Page 486
3.2 Decidability of the Invariance Problem......Page 488
4.2 Locations......Page 489
5.1 Summary......Page 490
5.3 Future Work......Page 491
References......Page 492
1 Introduction......Page 494
2 Preliminaries on Order-Sorted Algebra......Page 497
3 Order-Sorted Rewriting and Equality......Page 500
3.1 Kind-Complete OS-Rewriting and Equational Deduction......Page 501
3.2 Conservativity Results......Page 504
4.1 Abstract Congruence Closure......Page 505
4.2 Deciding OS (,)-QF-Satisfiability......Page 506
5 Order-Sorted (,AC)-QF-Satisfiability......Page 507
6 Related Work and Conclusions......Page 508
References......Page 509
Monads......Page 511
1 Introduction......Page 512
2 Parametric Monads Are Graded Monads......Page 516
3 Adjunction Pairs Induced from Graded Monads......Page 517
3.1 The Eilenberg-Moore Construction......Page 518
3.2 The Kleisli Construction......Page 519
3.3 Resolutions of Graded Monads......Page 520
4.1 A 2-Category for Eilenberg-Moore Objects......Page 521
4.2 A 2-Category for Kleisli Objects......Page 523
5 Illustration: The Graded State Monad......Page 525
References......Page 528
1 Introduction......Page 530
2 Preliminaries......Page 532
3 Profinite Monads......Page 536
4.1 Pseudovarieties and Profinite (In-)equations......Page 538
4.2 Reiterman's Theorem for T-algebras......Page 539
5 Applications and Examples......Page 542
6 Conclusions and Future Work......Page 544
References......Page 545
Author Index......Page 547
Recommend Papers

Foundations of Software Science and Computation Structures, 19 conf., FOSSACS 2016
 9783662496299, 9783662496305

  • 0 0 0
  • Like this paper and download? You can publish your own PDF file online for free in a few minutes! Sign Up
File loading please wait...
Citation preview

LNCS 9634

ARCoSS

Bart Jacobs Christof Löding (Eds.)

Foundations of Software Science and Computation Structures 19th International Conference, FOSSACS 2016 Held as Part of the European Joint Conferences on Theory and Practice of Software, ETAPS 2016 Eindhoven, The Netherlands, April 2–8, 2016, Proceedings

123

Lecture Notes in Computer Science

9634

Commenced Publication in 1973 Founding and Former Series Editors: Gerhard Goos, Juris Hartmanis, and Jan van Leeuwen

Editorial Board David Hutchison, UK Josef Kittler, UK Friedemann Mattern, Switzerland Moni Naor, Israel Bernhard Steffen, Germany Doug Tygar, USA

Takeo Kanade, USA Jon M. Kleinberg, USA John C. Mitchell, USA C. Pandu Rangan, India Demetri Terzopoulos, USA Gerhard Weikum, Germany

Advanced Research in Computing and Software Science Subline of Lecture Notes in Computer Science Subline Series Editors Giorgio Ausiello, University of Rome ‘La Sapienza’, Italy Vladimiro Sassone, University of Southampton, UK

Subline Advisory Board Susanne Albers, TU Munich, Germany Benjamin C. Pierce, University of Pennsylvania, USA Bernhard Steffen, University of Dortmund, Germany Deng Xiaotie, City University of Hong Kong Jeannette M.Wing, Microsoft Research, Redmond, WA, USA

More information about this series at http://www.springer.com/series/7407

Bart Jacobs Christof Löding (Eds.) •

Foundations of Software Science and Computation Structures 19th International Conference, FOSSACS 2016 Held as Part of the European Joint Conferences on Theory and Practice of Software, ETAPS 2016 Eindhoven, The Netherlands, April 2–8, 2016 Proceedings

123

Editors Bart Jacobs Radboud University Nijmegen Nijmegen The Netherlands

Christof Löding RWTH Aachen University Aachen Germany

ISSN 0302-9743 ISSN 1611-3349 (electronic) Lecture Notes in Computer Science ISBN 978-3-662-49629-9 ISBN 978-3-662-49630-5 (eBook) DOI 10.1007/978-3-662-49630-5 Library of Congress Control Number: 2016932745 LNCS Sublibrary: SL1 – Theoretical Computer Science and General Issues © Springer-Verlag Berlin Heidelberg 2016 This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. The publisher, the authors and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, express or implied, with respect to the material contained herein or for any errors or omissions that may have been made. Printed on acid-free paper This Springer imprint is published by Springer Nature The registered company is Springer-Verlag GmbH Berlin Heidelberg

ETAPS Foreword

Welcome to the proceedings of ETAPS 2016, which was held in Eindhoven, located in “the world’s smartest region,” also known as the Dutch Silicon Valley. Since ETAPS’ second edition held in Amsterdam (1999), ETAPS returned to The Netherlands this year. ETAPS 2016 was the 19th instance of the European Joint Conferences on Theory and Practice of Software. ETAPS is an annual federated conference established in 1998, consisting of five constituting conferences (ESOP, FASE, FoSSaCS, TACAS, and POST) this year. Each conference has its own Programme Committee and its own Steering Committee. The conferences cover various aspects of software systems, ranging from theoretical computer science to foundations to programming language developments, analysis tools, formal approaches to software engineering, and security. Organizing these conferences in a coherent, highly synchronized conference program, enables attendees to participate in an exciting event, having the possibility to meet many researchers working in different directions in the field, and to easily attend the talks of various conferences. Before and after the main conference, numerous satellite workshops took place and attracted many researchers from all over the globe. The ETAPS conferences received 474 submissions in total, 143 of which were accepted, yielding an overall acceptance rate of 30.2 %. I thank all authors for their interest in ETAPS, all reviewers for their peer-reviewing efforts, the Program Committee members for their contributions, and in particular the program co-chairs for their hard work in running this intensive process. Last but not least, my congratulations to all the authors of the accepted papers! ETAPS 2016 was greatly enriched by the unifying invited speakers Andrew Gordon (MSR Cambridge and University of Edinburgh, UK), and Rupak Majumdar (MPI Kaiserslautern, Germany), as well as the conference-specific invited speakers (ESOP) Cristina Lopes (University of California at Irvine, USA), (FASE) Oscar Nierstrasz (University of Bern, Switzerland), and (POST) Vitaly Shmatikov (University of Texas at Austin, USA). Invited tutorials were organised by Lenore Zuck (Chicago) and were provided by Grigore Rosu (University of Illinois at Urbana-Champaign, USA) on software verification and Peter Ryan (University of Luxembourg, Luxembourg) on security. My sincere thanks to all these speakers for their inspiring and interesting talks! ETAPS 2016 took place in Eindhoven, The Netherlands. It was organized by the Department of Computer Science of the Eindhoven University of Technology. It was further supported by the following associations and societies: ETAPS e.V., EATCS (European Association for Theoretical Computer Science), EAPLS (European Association for Programming Languages and Systems), and EASST (European Association of Software Science and Technology). The local organization team consisted of Mark van den Brand, Jan Friso Groote (general chair), Margje Mommers, Erik Scheffers, Julien Schmaltz, Erik de Vink, Anton Wijs, Tim Willemse, and Hans Zantema.

VI

ETAPS Foreword

The overall planning for ETAPS is the main responsibility of the Steering Committee, and in particular of its Executive Board. The ETAPS Steering Committee consists of an Executive Board and representatives of the individual ETAPS conferences, as well as representatives of EATCS, EAPLS, and EASST. The Executive Board consists of Gilles Barthe (Madrid), Holger Hermanns (Saarbrücken), Joost-Pieter Katoen (chair, Aachen and Twente), Gerald Lüttgen (Bamberg), Vladimiro Sassone (Southampton), and Tarmo Uustalu (Tallinn). Other members of the Steering Committee are: Parosh Abdulla (Uppsala), David Basin (Zurich), Giuseppe Castagna (Paris), Marsha Chechik (Toronto), Javier Esparza (Munich), Jan Friso Groote (Eindhoven), Reiko Heckel (Leicester), Marieke Huisman (Twente), Bart Jacobs (Nijmegen), Paul Klint (Amsterdam), Jens Knoop (Vienna), Kim G. Larsen (Aalborg), Axel Legay (Rennes), Christof Löding (Aachen), Matteo Maffei (Saarbrücken), Pasquale Malacaria (London), Tiziana Margaria (Limerick), Andrzej Murawski (Warwick), Catuscia Palamidessi (Palaiseau), Frank Piessens (Leuven), Jean-Francois Raskin (Brussels), Mark Ryan (Birmingham), Julia Rubin (Massachussetts), Don Sannella (Edinburgh), Perdita Stevens (Edinburgh), Gabriele Taentzer (Marburg), Peter Thiemann (Freiburg), Luca Vigano (London), Igor Walukiewicz (Bordeaux), Andrzej Wąsowski (Copenhagen), and Hongseok Yang (Oxford). I sincerely thank all ETAPS Steering Committee members for all their work in making the 19th edition of ETAPS a success. Moreover, thanks to all speakers, attendees, organizers of the satellite workshops, and Springer for their support. Finally, a big thanks to Jan Friso and his local organization team for all their enormous efforts enabling ETAPS to take place in Eindhoven! January 2016

Joost-Pieter Katoen ETAPS SC Chair ETAPS e.V. President

Preface

This volume contains the papers presented at the 19th International Conference on Foundations of Software Science and Computation Structures (FoSSaCS 2016) held in Eindhoven, The Netherlands, during April 4–7, 2016, as part of the European Joint Conferences on Theory and Practice of Software (ETAPS). The conference is dedicated to foundational research with a clear significance for software science and brings together research on theories and methods to support the analysis, integration, synthesis, transformation, and verification of programs and software systems. The 24 members of the Program Committee (PC) selected 31 papers out of 85 submissions. Each submission was reviewed by at least three PC members, with the help of external experts. After a three-day rebuttal phase, the selection was made based on discussions via the EasyChair conference management system. We wish to thank all authors who submitted to FoSSaCS 2016, all the PC members for their excellent work, and the external reviewers for their thorough evaluation of the submissions. In addition, we would like to thank the ETAPS organization for providing an excellent environment for FoSSaCS and other conferences and workshops. January 2016

Christof Löding Bart Jacobs

Organization

Program Committee Achim Blumensath Thomas Brihaye Arnaud Carayol Stéphane Demri Maribel Fernandez Nate Foster Marco Gaboardi Masahito Hasegawa Chris Heunen Jan Hoffmann Bart Jacobs Radha Jagadeesan Bartek Klin Naoki Kobayashi Manfred Kufleitner Orna Kupferman Paul Levy Christof Löding Matteo Mio Sylvain Salvati Olivier Serre Colin Stirling Nikos Tzevelekos Daniele Varacca

TU Darmstadt, Germany Université de Mons, France Université Paris Est, France ENS Cachan, France King’s College London, UK Cornell University, USA University of Dundee, UK Kyoto University, Japan University of Oxford, UK Carnegie Mellon University, USA Radboud University Nijmegen, The Netherlands DePaul University, USA University of Warsaw, Poland University of Tokyo, Japan University of Stuttgart, Germany Hebrew University of Jerusalem, Israel University of Birmingham, UK RWTH Aachen University, Germany ENS Lyon, France Université de Bordeaux, France Université Paris Diderot, France The University of Edinburgh, UK Queen Mary University of London, UK Université Paris Est Créteil, France

Contents

Types Comprehensive Parametric Polymorphism: Categorical Models and Type Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Neil Ghani, Fredrik Nordvall Forsberg, and Alex Simpson

3

Guarded Dependent Type Theory with Coinductive Types . . . . . . . . . . . . . . Aleš Bizjak, Hans Bugge Grathwohl, Ranald Clouston, Rasmus E. Møgelberg, and Lars Birkedal

20

Dependent Types and Fibred Computational Effects . . . . . . . . . . . . . . . . . . Danel Ahman, Neil Ghani, and Gordon D. Plotkin

36

Game Semantics for Bounded Polymorphism . . . . . . . . . . . . . . . . . . . . . . . James Laird

55

Recursion and Fixed-Points Join Inverse Categories as Models of Reversible Recursion . . . . . . . . . . . . . Holger Bock Axelsen and Robin Kaarsgaard

73

A Coalgebraic View of Bar Recursion and Bar Induction. . . . . . . . . . . . . . . Venanzio Capretta and Tarmo Uustalu

91

A New Foundation for Finitary Corecursion: The Locally Finite Fixpoint and Its Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Stefan Milius, Dirk Pattinson, and Thorsten Wißmann Fixed-Point Elimination in the Intuitionistic Propositional Calculus . . . . . . . . Silvio Ghilardi, Maria João Gouveia, and Luigi Santocanale

107 126

Verification and Program Analysis A Theory of Monitors: (Extended Abstract) . . . . . . . . . . . . . . . . . . . . . . . . Adrian Francalanza

145

Contextual Approximation and Higher-Order Procedures . . . . . . . . . . . . . . . Ranko Lazić and Andrzej S. Murawski

162

A Theory of Slicing for Probabilistic Control Flow Graphs . . . . . . . . . . . . . Torben Amtoft and Anindya Banerjee

180

X

Contents

Verification of Parameterized Communicating Automata via Split-Width . . . . Marie Fortin and Paul Gastin

197

Automata, Logic, Games Robust Equilibria in Mean-Payoff Games. . . . . . . . . . . . . . . . . . . . . . . . . . Romain Brenguier

217

Quantifier Alternation for Infinite Words . . . . . . . . . . . . . . . . . . . . . . . . . . Théo Pierron, Thomas Place, and Marc Zeitoun

234

Synchronizing Automata over Nested Words . . . . . . . . . . . . . . . . . . . . . . . Dmitry Chistikov, Pavel Martyugin, and Mahsa Shirmohammadi

252

On Freeze LTL with Ordered Attributes. . . . . . . . . . . . . . . . . . . . . . . . . . . Normann Decker and Daniel Thoma

269

Regular Transformations of Data Words Through Origin Information . . . . . . Antoine Durand-Gasselin and Peter Habermehl

285

Probabilistic and Timed Systems Trace Refinement in Labelled Markov Decision Processes . . . . . . . . . . . . . . Nathanaël Fijalkow, Stefan Kiefer, and Mahsa Shirmohammadi

303

Qualitative Analysis of VASS-Induced MDPs. . . . . . . . . . . . . . . . . . . . . . . Parosh Aziz Abdulla, Radu Ciobanu, Richard Mayr, Arnaud Sangnier, and Jeremy Sproston

319

Metric Temporal Logic with Counting . . . . . . . . . . . . . . . . . . . . . . . . . . . . Shankara Narayanan Krishna, Khushraj Madnani, and Paritosh K. Pandya

335

Distributed Synthesis in Continuous Time . . . . . . . . . . . . . . . . . . . . . . . . . Holger Hermanns, Jan Krčál, and Steen Vester

353

Proof Theory and Lambda Calculus Unary Resolution: Characterizing PTIME . . . . . . . . . . . . . . . . . . . . . . . . . . . Clément Aubert, Marc Bagnol, and Thomas Seiller

373

Focused and Synthetic Nested Sequents . . . . . . . . . . . . . . . . . . . . . . . . . . . Kaustuv Chaudhuri, Sonia Marin, and Lutz Straßburger

390

Strong Normalizability as a Finiteness Structure via the Taylor Expansion of k-terms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Michele Pagani, Christine Tasson, and Lionel Vaux

408

Contents

Reasoning About Call-by-need by Means of Types . . . . . . . . . . . . . . . . . . . Delia Kesner

XI

424

Algorithms for Infinite Systems Coverability Trees for Petri Nets with Unordered Data. . . . . . . . . . . . . . . . . Piotr Hofman, Sławomir Lasota, Ranko Lazić, Jérôme Leroux, Sylvain Schmitz, and Patrick Totzke

445

Shortest Paths in One-Counter Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . Dmitry Chistikov, Wojciech Czerwiński, Piotr Hofman, Michał Pilipczuk, and Michael Wehar

462

The Invariance Problem for Matrix Semigroups . . . . . . . . . . . . . . . . . . . . . Klaus Dräger

479

Order-Sorted Rewriting and Congruence Closure . . . . . . . . . . . . . . . . . . . . José Meseguer

493

Monads Towards a Formal Theory of Graded Monads. . . . . . . . . . . . . . . . . . . . . . . Soichiro Fujii, Shin-ya Katsumata, and Paul-André Melliès

513

Profinite Monads, Profinite Equations, and Reiterman’s Theorem . . . . . . . . . Liang-Ting Chen, Jiří Adámek, Stefan Milius, and Henning Urbat

531

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

549

Types

Comprehensive Parametric Polymorphism: Categorical Models and Type Theory Neil Ghani1 , Fredrik Nordvall Forsberg1(B) , and Alex Simpson2 1

2

Department of Computer and Information Sciences, University of Strathclyde, Glasgow, UK {neil.ghani,fredrik.nordvall-forsberg}@strath.ac.uk Faculty of Mathematics and Physics, University of Ljubljana, Ljubljana, Slovenia [email protected]

Abstract. This paper combines reflexive-graph-category structure for relational parametricity with fibrational models of impredicative polymorphism. To achieve this, we modify the definition of fibrational model of impredicative polymorphism by adding one further ingredient to the structure: comprehension in the sense of Lawvere. Our main result is that such comprehensive models, once further endowed with reflexive-graph-category structure, enjoy the expected consequences of parametricity. This is proved using a type-theoretic presentation of the category-theoretic structure, within which the desired consequences of parametricity are derived. The formalisation requires new techniques because equality relations are not available, and standard arguments that exploit equality need to be reworked.

1

Introduction

According to Strachey [26], a polymorphic program is parametric if it applies the same uniform algorithm at all instantiations of its type parameters. Reynolds [23] proposed relational parametricity as a mathematical model of parametric polymorphism. Relational parametricity is a powerful mathematical tool with many useful consequences; see [13, 21, 27] for numerous examples. The polymorphic lambda-calculus, λ2, (a.k.a. System F) was introduced independently by Girard [11] and Reynolds [22]. It serves as a model type theory for (impredicative) polymorphism, and thus provides a significant testing ground for ideas on relational parametricity. In this paper we address the question: What is the fundamental category-theoretic structure needed to model relational parametricity for λ2, which is both (i) minimal, in assuming as little structure as possible; but (ii) strong enough to ensure the expected consequences of parametricity hold? It is perhaps surprising that this question does not yet have an established answer. On the one hand, category-theoretic models for λ2 were developed many years ago by Seely [25]. They are studied systematically as λ2 fibrations in Jacobs [15]. On the other, the fundamental category-theoretic structure c Springer-Verlag Berlin Heidelberg 2016  B. Jacobs and C. L¨ oding (Eds.): FOSSACS 2016, LNCS 9634, pp. 3–19, 2016. DOI: 10.1007/978-3-662-49630-5 1

4

N. Ghani et al.

needed to model relational parametricity is also known. The crucial ingredient is the notion of reflexive graph category which appeared implicitly in Ma and Reynolds [19], was used explicitly by O’Hearn and Tennent [20], and Robinson and Rosolini [24], and reached maturity in the parametricity graphs of Dunphy and Reddy [7,8]. To obtain minimal structure for relational parametricity for λ2, it is natural to combine the structure of λ2 fibrations with that of parametricity graphs. This results in the notion of λ2 parametricity graph, which we define in Sect. 3. Sadly, λ2 parametricity graphs enjoy the expected properties of parametricity only in the special case that the underlying category is well-pointed. (Similar observations, for different but related notions of model, are made in [6–8].) Since well-pointedness rules out many categories of interest in semantics (e.g., functor categories) this limits the generality of the theory. One way of circumventing the restriction to well-pointed categories was proposed by Birkedal and Møgelberg [6], who developed a more elaborate categorytheoretic structure, which overcomes the limitation by modelling Plotkin and Abadi’s logic for parametricity [21]. This method of modelling the combination of λ2 with an extraneous logic has been refined and simplified by Hermida [12]. Nonetheless, it does not enjoy the simplicity in conception of combining the structure of category-theoretic models of λ2 with that of parametricity graphs. To obtain our minimal structure, we retain the original idea of combining parametricity graphs with category-theoretic models of λ2. However, we implement this in a perhaps unexpected way. We modify the notion of λ2 model. We ask for λ2 fibrations to additionally satisfy Lawvere’s comprehension property. Not only are the resulting comprehensive λ2 fibrations natural in their own right as models of λ2, but, when combined with parametricity-graph structure to form comprehensive λ2 parametricity graphs, they do indeed enjoy all expected consequences of parametricity. Sections 2 and 3 define comprehensive λ2 fibrations and comprehensive λ2 parametricity graphs respectively. In Sect. 4, we present a type theory λ2R, corresponding to our category-theoretic structure, which provides a simple system for reasoning about parametricity. The type theory λ2R is similar to Dunphy’s System P [7], and Abadi, Cardelli and Curien’s System R [1], to which it is compared in Sect. 7. In Sect. 5, we develop the technical machinery needed to reason in λ2R. A key obstacle is that the system does not include equality relations. This means that graph relations, which are a crucial ingredient in standard arguments involving relational parametricity, are not in general definable. In Sect. 5, we instead identify two forms of pseudograph relations, whose subtle interrelationship allows us to establish the consequences we need. One kind of pseudograph relation is immediately definable using the fibrational structure built into the notion of parametricity graph. The other type of pseudograph requires opfibrational structure. We use an impredicative encoding to show that opfibrational structure is definable in λ2R, and hence always present in comprehensive λ2 parametricity graphs. In Sect. 6, we finally apply the technical machinery and establish that the expected consequences of relational parametricity are indeed derivable in λ2R, and hence hold in comprehensive λ2 parametricity graphs.

Comprehensive Parametric Polymorphism: Categorical Models

5

In summary, the main contributions of this work are: (i) The definition of comprehensive λ2 fibrations as models of λ2. (ii) The definition of a new category-theoretic notion of model of relational parametricity, obtained by combining parametricity graphs and comprehensive λ2 fibrations into comprehensive λ2 parametricity graphs. (iii) The extraction of λ2R as the type theory intrinsic to comprehensive λ2 parametricity graphs. (iv) The derivation of the expected consequences of parametricity in λ2R, and hence in comprehensive λ2 parametricity graphs. This requires novel techniques: establishing the opfibration property of comprehensive λ2 parametricity graphs, and the use of pseudograph relations. In the category-theoretic parts of the paper, we assume familiarity with fibred category theory, for which Jacobs [15] is our main reference. Nevertheless, a substantial portion of the paper is presented in purely type-theoretic terms, and may be read without reference to the accompanying category-theoretic material.

2

Comprehensive λ2 Fibrations

In Fig. 1, we recall the polymorphic λ-calculus λ2. We use x, y, . . . to range over term variables, and α, β, . . . to range over type variables. Our presentation has four judgements: Γ ctxt , stating that Γ is a well-formed context; Γ  A type , stating that A is a well-formed type in context Γ ; Γ  t : A , stating that the term t has type A in context Γ ; and judgemental equality Γ  t1 = t2 : A . We assume β and η-equalities for both term abstraction, λ, and type abstraction, Λ. Equality is also assumed to be a congruence relation, although the rules guaranteeing this have been omitted from Fig. 1 for brevity. A minor departure from many presentations of λ2 is that we interleave type variables and term variables in a single context. This approach is not only natural, but indeed standard when λ2 is considered in the context of dependent type theory; for example, when derived as an instance of a pure type system [3]. Since there is no dependency of λ2 types on term variables, such interleaving is syntactically vacuous. Nevertheless, we shall see below that its presence does have semantic implications. We next recall the standard category-theoretic notion of λ2 fibration, which models λ2. We directly restrict the definition to the split case to circumvent coherence issues that would otherwise arise, cf. [15]. Definition 1 (λ2 fibration). A λ2 fibration is a split fibration p : T → C, where the base category C has finite products, and the fibration: (i) is fibred cartesian closed; (ii) has a split generic object U [15, Definition 5.2.1] — we write Ω for p U ; (iii) and has fibred-products along projections X × Ω - X in C.

6

N. Ghani et al. Context formation rules: Γ ctxt Γ A type (x ∈ / Γ) Γ, x : A ctxt

Γ ctxt (α ∈ / Γ) Γ, α ctxt

· ctxt

Type formation rules:

Γ

α type

(α ∈ Γ )

Γ

A type Γ B type Γ A → B type

Γ, α A type Γ α. A type

Term typing rules: Γ

x:A

(x : A ∈ Γ )

Γ

Γ

Γ, x : A t : B λx. t : A → B

Γ, α t : A Λα. t : ∀α. A

Γ

Γ

s:A→B Γ Γ st : B

t:A

t : ∀α. A Γ B type Γ t[B] : A[α → B]

Judgemental equality: Γ

Γ

(λx. t)u = t[x → u] : B

(Λα. t)[U ] = t[α → U ] : A[α → U ]

Γ

t = λx. tx : A → B

Γ

(x ∈ / Γ)

t = Λα. t[α] : ∀α. A

(α ∈ / Γ)

Fig. 1. The type system λ2

Moreover, the reindexing functors given by the splitting are required to preserve the above-specified structure in fibres on the nose. The above definition differs slightly from [15, Definition 8.4.3(b)] in that we do not include fibred coproducts in condition (iii). These are not needed to model λ2, and their existence is anyway derivable in parametric models. In a λ2 fibration, we write TX for the fibre category over X. We also use X as a subscript when referring to structure in TX ; e.g., 1X is the specified terminal - Y object in TX , and ⇒X is the exponential structure in TX . Given f : X ∗ ∗ ∗ in C, we write f for the reindexing functor TY → TX , and A f : fA - A for the specified cartesian lifting of f relative to A. We also write Ω for the specified right adjoint, given by (iii), to reindexing functors π1∗ : TX → TX×Ω . We recall in outline the semantic interpretation of λ2 in a λ2 fibration T → C. A context Θ = α1 , . . . , αn of type variables is interpreted as the n-fold product Θ = Ω n in C. A type A in type-variable context Θ is then interpreted as an object AΘ of T over Θ, defined by induction on the structure of A, using cartesian closure for function types, fibred products for universal types, and the - Ω reindexing (πi )∗ U of the generic object along the projection πi : Ω n to interpret αi over Θ. Finally, the interpretation of a term Γ  t : A is

Comprehensive Parametric Polymorphism: Categorical Models

7

obtained by splitting Γ into its component contexts: Θ of type variables, and Δ of term variables. Then Δ = x1 : A1 , . . . , xm : Am is interpreted as the product ΔΘ = A1 Θ × · · · × Am Θ in the fibre over Θ, and t is interpreted as a morphism tΓ : ΔΘ - AΘ in TΘ . In the above outline, one sees that the structure of a λ2 fibration fits uneasily alongside our mixed contexts of interleaved type and term variables, since these have to be separated to define the semantic interpretation. In dependent type theory, where no such separation is possible, a more direct semantic interpretation is achieved using Lawvere’s comprehension property [18] to model the process of context extension [14]. It is natural to apply the same idea to λ2. Definition 2 (Comprehensive λ2 fibration). A λ2 fibration p : T → C is comprehensive if it enjoys the comprehension property [15, Definition 10.4.7]: the terminal-object functor X → 1X : C → T has a specified right adjoint K : T → C. Requiring a specified right adjoint maintains consistency with our policy of working with split fibrational structure. Given A in TX , we write κA : KA - X for the ‘projection’ map obtained by applying p to the counit 1KA - A in T. To show that comprehensive λ2 fibrations permit a direct, inductive-onsyntax semantic interpretation, we present the interpretation of λ2 types in detail. A context Γ ctxt is interpreted as an object Γ  of C; and a type Γ  A type is interpreted as an object AΓ in TΓ  . These are defined by mutual induction, together with maps πΓα : Γ  - Ω for every context Γ containing α. · = 1

αΓ = (πΓα )∗ U

Γ, α = Γ  × Ω A → BΓ = AΓ ⇒Γ  BΓ  ∀α. A = AΓ, α Γ, x : A = KAΓ Ω

α πΓ, α = π2 α α πΓ, β = πΓ ◦ π1 (β = α) α α πΓ, x:A = πΓ ◦ κAΓ

Having made the above definitions, a term Γ  t : A is interpreted as a global element tΓ : 1Γ  - AΓ in TΓ  . The definition, which we omit, is a straightforward induction on the derivation of Γ  t : A. The appropriateness of comprehensive λ2 fibrations as a notion of model for λ2 is supported by soundness and completeness results. Theorem 3 (Soundness for λ2). If Γ  t1 = t2 : A then, in every comprehensive λ2 fibration, we have t1 Γ = t2 Γ . Theorem 4 (Full completeness for λ2). There exists a comprehensive λ2 fibration satisfying: (i) for every type Γ  A type, every global point 1Γ  - AΓ is the denotation tΓ of some term Γ  t : A; and (ii) for all terms Γ  t1 , t2 : A satisfying t1 Γ = t2 Γ , we have Γ  t1 = t2 : A. Theorem 3 is proved by a routine induction on equality derivations, and Theorem 4 by construction of a syntactic model, which has the requisite properties.

8

3

N. Ghani et al.

Comprehensive λ2 Parametricity Graphs

Reflexive graph categories are studied in [7, 8,16, 19, 24] as a simple categorytheoretic structure for modelling relational parametricity. A reflexive graph category consists of a pair of categories, V, the vertex category, and E, the edge category, together with functors ∇1 , ∇2 : E → V and Δ : V → E satisfying ∇1 Δ = idV = ∇2 Δ. Informally, one thinks of E as a category whose objects are binary ‘relations’ between objects of V. Then ∇1 , ∇2 are ‘projection’ functors, and Δ maps an object to its ‘identity relation’. We shall be guided by the following general thesis. A model of relational parametricity, irrespective of the type theory for which it is considered, should form a reflexive graph category, in the (2-)category of structure-preserving functors between models of the type theory in question. This thesis is supported by the following considerations. Endowing the edge category E with the categorical structure needed to interpret types corresponds to giving types a relational interpretation. The preservation of this structure by the projection functors ∇1 , ∇2 means that the relational interpretation commutes with the usual non-relational interpretation of types. The preservation of structure by Δ, in combination with the identity property discussed later, corresponds to Reynolds’ identity extension property [23]. In the context of the present paper, we need to specialise the above recipe to (comprehensive) λ2 fibrations. A morphism from one (comprehensive) λ2 fibration p : T → C to another p : T → C is given by a pair of functors, H : T → T and L : C → C such that p H = L p , and such that H, L preserve all other specified structure (including the choice of cartesian morphisms in the splitting) on the nose. By a reflexive graph of (comprehensive) λ2 fibrations, we thus mean a pair of (comprehensive) λ2 fibrations with functors between them: ∇T1 , ΔT , ∇T2  - T R(T) pR

p

(1)

? - ?  - C R(C) C C ∇C 1 , Δ , ∇2 T C T C where each of the three pairs ∇T1 , ∇C 1 and ∇2 , ∇2 and Δ , Δ is a morphism of (comprehensive) λ2 fibrations, and where each of the triples ∇T1 , ∇T2 , ΔT and C C R ∇C 1 , ∇2 , Δ is a reflexive graph category. We emphasise that p : R(T) → R(C), in (1), is an arbitrary (comprehensive) λ2 fibration fitting into the diagram. The notation R(·) is merely mnemonic, and does not imply that R(T) is obtained using a particular construction from T. One needs to add further conditions to the above structure to ensure that the objects of R(T) behave sufficiently like relations. In [19], this was addressed by requiring the fibre category R(T)1R(C) , over the terminal object, to coincide with

Comprehensive Parametric Polymorphism: Categorical Models

9

a particular category of logical relations over T1C . As well as only being applicable if T1C has (sufficient) finite limits, this requirement also has the weakness that it says nothing about other fibres of pR . As a result, the structure is too weak to imply consequences of parametricity in general, see [6, 24] for discussion. To remedy this, we instead need axiomatic structure for a category of relations, in a form that is suitable for being imposed fibrewise on pR . This is provided by Dunphy and Reddy’s notion of parametricity graph [7, 8], which we now recall. A reflexive graph category ∇1 , ∇2 : E → V, Δ : V → E is said to be relational if the functor ∇1 , ∇2 : E → V × V is faithful. This property allows one to think of morphisms in E as pairs of relation-preserving maps from V. Accordingly, we call objects of E relations, we write R : A ↔ B to mean an object R of - S to mean E with ∇1 R = A and ∇2 R = B, and we write f × g : R - S in E with ∇1 h = f and that there is a (necessarily unique) map h : R ∇2 h = g. A reflexive graph category satisfies the identity property if, for every h : ΔA - ΔB in E, it holds that ∇1 h = ∇2 h. This allows one to think of ΔA as an identity relation on A (although, cf. Sect. 5 for caveats). In a relational reflexive graph category, the identity property is equivalent to the fullness of the functor Δ. A parametricity graph is a relational reflexive graph category with the identity property, for which the functor ∇1 , ∇2 : E → V × V is a fibration. The fibration property supports the following definition mechanism. Let R : A ↔ B - A and g : B  - B in be a relation in E. Then, given morphisms f : A V, reindexing produces an inverse image relation [f ×g]−1 R : A ↔ B  .1 The main category-theoretic definition of this paper is a fibrewise adaptation of parametricity graph to the context of comprehensive λ2 fibrations. Definition 5 ((Comprehensive) λ2 parametricity graph). A (comprehensive) λ2 parametricity graph is a reflexive graph of (comprehensive) λ2 fibrations, as in (1), that satisfies, for all objects W of R(C) and X of C: (Relational) The functor ∇T1 , ∇T2 R(T)W : R(T)W → T∇C1 W ×T∇C2 W is faithful. (Identity property) The functor ΔT TX : TX → R(T)ΔC X is full. (Fibration) ∇T1 , ∇T2 R(T)W : R(T)W → T∇C1 W ×T∇C2 W is a cloven fibration. Moreover, for every φ : W  R(T)W

∇T1 , ∇T2 R(T)W ? T∇C1 W ×T∇C2 W

- W in R(C), we require the commuting square (pR )∗ φ

- R(T)W 

∇T1 , ∇T2 R(T)W 

? - T∇C W  ×T∇C W  1 2 ∗ C p∗ (∇C 1 φ) × p (∇2 φ)

(where the notation distinguishes reindexing functors determined by p and pR ) to give a cleavage-preserving fibred functor from ∇T1 ,∇T2 R(T)W to

∇T1 ,∇T2 R(T)W  . 1

We use (·)−1 rather than (·)∗ for reindexing to emphasise that we are in a relational setting: ∇1 , ∇2  is a preorder fibration since it is faithful.

10

N. Ghani et al.

This definition could by strengthened by asking for the parametricity-graph fibrations to be split instead of merely cloven. Such a strengthening does not affect any of the results in the sequel, and may seem natural given our use of split fibrations in all previous definitions. Nevertheless, our choice of definition reflects the fact that the weaker cloven assumption is all that is needed to avoid coherence issues arising in the semantic interpretation of the type theory λ2R introduced in Sect. 4 below. It is Definition 5, with the comprehension property included, that provides our answer to the question highlighted in the introduction. (The definition without comprehension is included for comparison purposes only.)

4

A Type System for Relational Reasoning

We define a type system λ2R, suggested by the structure of comprehensive λ2 parametricity graphs. This system is similar, in many respects, to System R of Abadi, Cardelli and Curien [1] and System P of Dunphy [7], to which we shall compare it in Sect. 7. The rules for λ2R are given by Fig. 1 (it extends λ2) in combination with Fig. 2. The latter adds three new judgements: Θ rctxt says that Θ is a welldefined relational context; Θ  A1 RA2 rel says that R is a relation between types A1 and A2 , in relational context Θ; and Θ  (t1 : A1 )R(t2 : A2 ) is a relatedness judgement, asserting that t1 : A1 is related to t2 : A2 by the relation R. Relations, in Fig. 2, are built up from a collection of relation variables ρ, . . . , which, for clarity, we choose to keep disjoint from type and term variables. In the rules, we make use of three operations (·)1 , (·)2 and · , defined in Fig. 3, which implement reflexive graph structure on syntax. The (·)i operations project a relational context to a typing context, whereas the · operation acts in the other direction. In the definition of the latter, we associate a distinct relation variable ρα to every type variable α. Lemma 7 below states how these operations relate typing and relational judgements. The rules for building relational contexts and relations, in Fig. 2, require some explanation. In adding an assertion αρβ to a relational context Θ, all variables α, β, ρ need to be sufficiently fresh. However, the formulation of λ2R is such that variables on the left-hand side of relations are always manipulated separately from variables on the right. Thus, for example, α is sufficiently fresh in αρβ, as long as α does not already occur on the left side (Θ)1 of Θ. A similar separation principle applies also with respect to the term variables x1 , x2 in assertions (x1 : A1 )R(x2 : A2 ). The separation principle means that one needs to be cautious in interpreting assertions of the form αρα and (x : A)R(x : A). In such assertions, even though the same variable appears on the left and right, the correct intuition is that these are really two distinct variables. We have chosen not to underline this distinction by requiring the variables to be syntactically different, since doing so would add unnecessary syntactic clutter to the system; for example, it would complicate the definition of the · operation. Instead, we

Comprehensive Parametric Polymorphism: Categorical Models

Relational context formation rules: Θ rctxt (ρ ∈ / Θ, α ∈ / Θ1 , β ∈ / Θ2 ) Θ, αρβ rctxt

· rctxt

Θ rctxt Θ A1 RA2 rel / Θ1 , x2 ∈ / Θ2 ) (x1 ∈ Θ, (x1 : A1 )R(x2 : A2 ) rctxt Relation formation rules: Θ

αρβ rel

Θ Θ

Θ

(αρβ ∈ Θ)

Θ

Θ B1 SB2 rel A1 RA2 rel (A1 → B1 )(R → S)(A2 → B2 ) rel

Θ, αρβ A1 RA2 rel (∀α. A1 )(∀αρβ. R)(∀β. A2 ) rel

B1 RB2 rel Θ

Θ1 t1 : A1 → B1 Θ2 A1 ([t1 × t2 ]−1 R)A2 rel

t2 : A2 → B2

Relatedness rules: Θ

(x1 : A1 )R(x2 : A2 )

Θ, (x1 : A1 )R(x2 : A2 ) (t1 : B1 )S(t2 : B2 ) (λx1 . t1 : A1 → B1 )(R → S)(λx2 . t2 : A2 → B2 )

Θ

(s1 : A1 → B1 )(R → S)(s2 : A2 → B2 ) Θ Θ (s1 t1 : B1 )S(s2 t2 : B2 )

Θ

Θ Θ Θ

Θ

(Λα. t1 : ∀α. A1 ) ∀αρβ. R (Λβ. t2 : ∀β. A2 )

(t1 : ∀α. A1 ) ∀αρβ. R (t2 : ∀β. A2 )

Θ

B1 SB2 rel

(t1 [B1 ] : A1 [α → B1 ])R[αρβ → B1 SB2 ](t2 [B2 ] : A2 [β → B2 ])

(t1 u1 : B1 )R(t2 u2 : B2 )

(u1 : A1 )([t1 × t2 ] Θ

(t1 : A1 )R(t2 : A2 )

(t1 : A1 )R(t2 : A2 )

Θ, αρβ

Θ

((x1 : A1 )R(x2 : A2 ) ∈ Θ)

−1

Θ

R)(u2 : A2 )

(t1 : A1 )R(t2 : A2 ) Θ

(u1 : A1 )([t1 × t2 ]−1 R)(u2 : A2 ) Θ (t1 u1 : B1 )R(t2 u2 : B2 )

Θ1 t1 = s1 : A1 (s1 : A1 )R(s2 : A2 )

Parametricity rule: Γ Γ

(s : A) A (t : A) s=t:A

Fig. 2. The type system λ2R

Θ2

t2 = s2 : A2

11

12

N. Ghani et al.

The operations (−)i (for i ∈ {0, 1}) on relational contexts: (·)i = · (Θ, α1 ρα2 )i = (Θ)i , αi (Θ, (x1 : A1 )R(x2 : A2 ))i = (Θ)i , xi : Ai The operation

on contexts and types: α = ρα

= Γ, α = Γ , α ρα α

A→B =( A

Γ, x : A = Γ , (x : A) A (x : A)

B )

α. A = ∀α ρα α. A

Fig. 3. Syntactic reflexive graph structure

rely on left and right positioning to make the necessary distinctions. This is crucial in the definition of the substitution operations on relations. There are two such operations: R[αρβ → ASB] substitutes, in the relation R, the type A for all left occurrences of α, the type B for all right occurrences of β (which may itself be α), and the relation S for all occurrences of ρ; similarly, S[x → s, y → t] substitutes, in the relation S, the term s for all left occurrences of x, and the term t for all right occurrences of y (which may itself be x). Note that relations can indeed contain terms and (hence) type variables, due to the [t1 × t2 ]−1 R construction, where we consider t1 as occurring on the left, and t2 on the right. Lemma 6 (Substitution lemma) (i) If Θ  A1 RA2 rel and Θ, α1 ρ α2  (t1 : B1 )S(t2 : B2 ) then Θ  (t1 [α1 → A1 ] : B1 [α1 → A1 ])S[α1 ρα2 → A1 RA2 ](t2 [α2 → A2 ] : B2 [α2 → A2 ]). (ii) If Θ  (t1 : A1 )R(t2 : A2 ) and Θ, (x1 : A1 )R(x2 : A2 )  (s1 : B1 )S(s2 : B2 ) then Θ  (s1 [x1 → t] : B1 )S[x1 → t1 , x2 → t2 ](s2 [x2 → t2 ] : B2 ). The relatedness rules of Fig. 2 include the expected rules for relations R → S and ∀αρβ. R, which mimic the analogous type constructions in λ2. The rules for [t1 × t2 ]−1 R implement its intended interpretation as an inverse image construction. In addition, a further rule expresses an extensionality principle for relations with respect to judgemental equality. Such an intermixing of relatedness judgements with equality judgements is legitimised by statement (i) of the lemma below. Lemma 7 (i) If Θ  (t1 : A1 )R(t2 : A2 ) then (Θ)i  ti : Ai . (ii) If Γ  t : A then Γ  (t : A) A (t : A). Statement (ii) of the lemma asserts that all terms enjoy the characteristic relation-preservation property of relational parametricity. By the extensionality

Comprehensive Parametric Polymorphism: Categorical Models

13

rule, it follows that Γ  s = t : A implies Γ  (s : A) A (t : A). That is, equal terms are parametrically related. Since parametric relatedness captures a form of behavioural equivalence, we can ask also for the converse implication to hold. This is implemented by the parametricity rule in Fig. 2. This rule, in the general form given, is derivable from its empty-context version:  (s : A) A (t : A) implies  s = t : A. Thus the parametricity rule is equivalent to asking for the relational interpretation of a closed type to act as an identity relation between closed terms—a weak version of Reynold’s identity extension property [23]. We discuss the relational interpretation of open types in Sect. 5. We outline the semantic interpretation of λ2R. Given a comprehensive λ2 parametricity graph, the contexts, types and terms of λ2 are interpreted in the comprehensive λ2 fibration p : T → C, as in Sect. 2. In addition, we interpret a relational context Θ as an object Θ of R(C), and a syntactic relation Θ  ARB rel as a semantic relation RΘ : A(Θ)1 ↔ B(Θ)2 in R(T)Θ . The definitions of Θ and RΘ interpret context extension, function space and universal quantification using the structure of the comprehensive λ2 fibration pR : R(T) → R(C), where relation variables αρβ are interpreted using the generic object of pR . For the inverse-image relation Θ  A1 ([t1 × t2 ]−1 R)A2 rel, - B1 (Θ) and we have that t1 (Θ)1 and t2 (Θ)2 determine maps A1 (Θ)1 1 - B2 (Θ) in T(Θ)  and T(Θ)  respectively. The fibration propA2 (Θ)2 2 1 2 erty of ∇T1 , ∇T2 R(T)Θ then gives [t1 × t2 ]−1 R : A1 (Θ)1 ↔ A2 (Θ)2 as the inverse image of R : B1 (Θ)1 ↔ B2 (Θ)2 along these maps. In the above semantic interpretation, the comprehension property is needed in order to interpret a relational context Θ as an object Θ of R(C), and essential use is made of this in the definition of [t1 ×t2 ]−1 R. Were the comprehension property of models dropped, it would be possible to rejig the semantics to interpret a restricted calculus with inverse-image relations definable only in relational contexts containing no term variables, but not full λ2R. The semantics is supported by soundness and completeness theorems. Theorem 8 (Soundness for λ2R). In every comprehensive λ2 parametricity graph: (i) if Γ  t1 = t2 : A then t1 Γ = t2 Γ ; and (ii) if Θ  (t1 : A1 )R(t2 : A2 ) then t1 (Θ)1 × t2 (Θ)2 : 1Θ

- RΘ .

Theorem 9 (Full completeness for λ2R). There exists a comprehensive λ2 parametricity graph satisfying the following. - AΓ is the deno(i) For every type Γ  A type, every global point 1Γ  tation tΓ of some term Γ  t : A. (ii) For all terms Γ  t1 , t2 : A satisfying t1 Γ = t2 Γ , we have Γ  t1 = t2 : A. - RΘ (iii) For every relation Θ  A1 RA2 type, every global point 1Θ arises as t1 (Θ)1 × t2 (Θ)2 for terms t1 , t2 such that Θ  (t1 : A1 )R(t2 : A2 ). Theorem 8 is proved by induction on derivations. We highlight that the soundness of the parametricity rule follows from the identity property of comprehensive λ2 parametricity graphs. Theorem 9 is proved by a term model construction.

14

5

N. Ghani et al.

Direct-Image and Pseudograph Relations

As already discussed, the parametricity rule of Fig. 2 interprets the relation

A as an identity relation when A is a closed type. When A contains type variables, however, this interpretation is not available. Consider an open type α  A(α) type (where we write A(α) to highlight the occurrences of α in A).  Then we have αρα  A(α) A (ρ) A(α) rel. However, the independent handling of left and right variables in λ2R (forced by the semantic correspondence with comprehensive λ2 parametricity   graphs), means that the latter relation is equivalent to αρβ  A(α) A (ρ) A(β) rel; i.e., it is a family (indexed by relations ρ) of relations between different types. Indeed, the distinctness of left and right type variables means λ2R has no facility for formulating relations between open types and themselves. In particular, λ2R contains no mechanism for defining identity relations on open types. Nonetheless, the relation A can act as a kind of pseudo-identity relation for type A where the parametricity rule can establish equalities from A -relatedness in relational contexts of the form Γ . Graphs of functions are ubiquitous in standard arguments involving relational parametricity. Since we have only pseudo-identity relations, we correspondingly have only pseudographs available in λ2R. Suppose Γ  f : A → B. Define: gr∗ (f ) := [f × idB ]−1 B Clearly Γ  A gr∗ (f ) B rel. Its defining property is that (x : A) gr∗ (f ) (y : B) holds if and only if (f x : B) B (y : B). Mathematically, there is, however, another natural pseudograph relation, for f , between A and B. This is the relation gr! (f ) defined by (x : A)gr! (f )(y : B) if there exists w : A such that (x : A) A (w : A) and y = f w. Since, by (ii) of Lemma 7, f maps A -related values to B -related values, gr! (f ) ⊆ gr∗ (f ). However, because A and B are not identity relations, there is no need for this inclusion to be an equality. We shall need to make use of both forms of pseudograph relation to derive the standard consequences of parametricity. In order to do so, we must first provide a definition of gr! (f ) in λ2R itself, and establish formal analogues of the informal observations above. The main construction we need is that of direct-image relations [t1 × t2 ]! R, dual to inverse-image relations. This is achieved using an impredicative encoding. Theorem 10 (Direct-image relations). Using the definition [t1 × t2 ]! R := [iB1 × iB2 ]−1 (∀αρα. ([(− ◦ t1 ) × (− ◦ t2 )]−1 (R → ρ)) → ρ) where iB abbreviates λb. Λα. λt. t b : B → ∀α. (B → α) → α and (−◦tj ) abbreviates λ(vj : Bj → α).λ(xj : Aj ).vj (t1 xj ), λ2R supports the derived rules below. Θ1  t1 : A1 → B1 Θ2  t2 : A2 → B2 Θ  A1 RA2 rel Θ  B1 ([t1 × t2 ]! R)B2 rel

Θ  (u1 : A1 )R(u2 : A2 )   Θ  (t1 u1 : B1 ) [t1 × t2 ]! R (t2 u2 : B2 )

Θ  C1 QC2 rel Θ1  v1 : B1 → C1  Θ2  v2 : B2 → C2   Θ  (u1 : B1 ) [t1 × t2 ]! R (u2 : B2 ) Θ  (v1 ◦ t1 : A1 → C1 ) R → Q (v2 ◦ t2 : A2 → C2 Θ  (v1 u1: C1 )Q(v2 u2 : C2 )

Comprehensive Parametric Polymorphism: Categorical Models

15

In fact, these rules are derivable without use of the parametricity rule of λ2R. It is now straightforward to define the second form of pseudograph relation discussed above. Suppose that Γ  f : A → B and define Γ  A gr! (f ) B rel by: gr! (f ) := [idA × f ]! A . To understand the relationship between the two pseudograph relations we introduce some notation. Given R and S such that Θ  ARB rel and Θ  ASB rel, let Θ  R ⊆ S abbreviate Θ, (x : A)R(y : B)  (x : A)S(y : B). Lemma 11. If Γ  f : A → B then: (i) Γ  gr! (f ) ⊆ gr∗ (f ); and (ii) Γ  (s : A) gr∗ (f ) (t : B) iff Γ  f s = t : B iff Γ  (s : A) gr! (f ) (t : B). We comment that, in spite of item (ii), the converse inclusion to (i) does not hold in general. Property (ii) applies only in context Γ , and thus implies nothing about what happens if further relational assumptions are added. Theorem 10 has a semantic analogue: direct image relations correspond to opfibrational structure on comprehensive λ2 parametricity graphs. Theorem 12. In any comprehensive λ2 parametricity graph, for every object W of R(C), the functor ∇T1 , ∇T2 R(T)W : R(T)W → T∇C1 W ×T∇C2 W is an opfibration.

6

Consequences of Parametricity

System λ2R is strong enough to prove the familiar consequences of parametricity. Theorem 13 (Consequences of Parametricity). System λ2R proves: The unit (terminal) type can be encoded as 1 = ∀α. α → α. The product of A and B can be encoded as A×B = ∀α. (A → B → α) → α. The empty (initial) type can be encoded as 0 = ∀α. α. The sum of A and B can be encoded as A+B = ∀α. (A → α) → (B → α) → α. Existential types can be encoded as ∃α. T (α) = ∀α. (∀β. (T (β) → α)) → α. The type ∀α. (T (α) → α) → α is the carrier of the initial T -algebra for all functorial type expressions T (α). (vii) The type ∃α. (α → T (α)) × α is the carrier of the final T -coalgebra for all functorial type expressions T (α). (i) (ii) (iii) (iv) (v) (vi)

This result for λ2R implies that analogous category-theoretic properties (which we do not state for lack of space) hold for comprehensive λ2 parametricity graphs. The proofs of (i)–(vii) follow the usual ones, see, e.g., [21], but with graph relations replaced by pseudographs. Pseudograph relations of the form gr∗ (f ) suffice in all proofs with the exception of the verification of final coalgebras, where

16

N. Ghani et al.

gr! (f ) is used. In this section, we explain how this difference in the treatment of initial algebras and final coalgebras arises. For lack of space, we focus on the use of pseudograph relations only, and omit the (standard) supporting arguments. Suppose Γ, α  T type. We write T (α) to highlight the occurrences of α in T , and T (A) for the substitution T [α → A]. If α occurs only positively in T (i.e., not on the left-hand side of an odd number of arrows) then it is standard that T defines an endofunctor on types. If Γ   f : A → B, where Γ  extends Γ , then we use the notation Γ   T (f ) : T (A) → T (B) for the functorial action of T . This action preserves identities and composition up to judgemental equality. In addition, the corresponding relational substitution preserves pseudo-identity relations; i.e., T ( A ) (by which we mean the substitution T [αρα α → A ]) syntactically coincides with T (A) . Also, the functorial action lifts to relations: if Θ  (f : A → B)(R → S)(f  : A → B  ), where Θ extends Γ , then: Θ  (T (f ) : T (A) → T (B))( T (R) → T (S))(T (f  ) : T (A ) → T (B  )). Using these facts (which assert that T is a reflexive-graph functor [8]) one establishes the following properties of the action of T on pseudograph relations. Lemma 14. Suppose α occurs positively in Γ, α  T type and Γ   f : A → B, where Γ  extends Γ . (i) Γ   T (gr∗ (f )) ⊆ gr∗ (T (f )) . (ii) Γ   gr! (T (f )) ⊆ T (gr! (f )) . Our proof of this lemma closely mirrors the proof of the Graph Lemma in [9], which exploits the fact that graph relations can be defined either using inverse image, analogously to gr∗ (f ), or using direct image, analogously to gr! (f ). We now explain how Lemma 14 bears on the proofs of the universal properties of initial algebras and final coalgebras. Given T as above, standard constructions produce a T -algebra and a T -coalgebra that can be shown to be weakly initial and weakly final respectively, without invoking parametricity. The parametricity rule is used to establish the uniqueness part of the universal property. In the initiality and finality arguments, one is led to consider T -algebra and T -coalgebra homomorphisms respectively: T (A)

T (h)-

a

? A

h

T (B)

b ? - B

T (A) 6 a A

T (h )-

h

T (B) 6 b

- B

where the diagrams are given by terms, in a context Γ  extending Γ , which commute up to judgemental equality. Lemma 14 allows one to prove the following crucial properties as consequences of the commutativity of the above diagrams.  

Γ   (a : T (A) → A) T (gr∗ (h)) → gr∗ (h) (b : T (B) → B)  

Γ   (a : A → T (A)) gr! (h ) → T (gr! (h )) (b : B → T (B))

Comprehensive Parametric Polymorphism: Categorical Models

17

It is the orientation of the function relations above that necessitates the use of a different type of pseudograph relation in each case. Modulo this subtlety, the remaining proofs of initiality and finality proceed as usual, cf. [21].

7

Related and Further Work

System R of [1] and System P of [7] share with λ2R the property of having a syntax in which function space and universal quantification are basic constructions on relations. Indeed λ2R is especially similar to System P, which also has the inverse-image-relation constructor [t1 × t2 ]−1 R. The most significant difference is that, in System P, the formation rule for this construction is restricted: the terms t1 , t2 are not allowed to contain free term variables. However, they are permitted to contain so-called indeterminates, which, in the semantics of System P, range over global elements in models. This device allows System P to be used to establish consequences of parametricity in well-pointed models [7]. In λ2R, our general arguments for consequences of parametricity make essential use of the possibility for t1 and t2 to contain free term variables. As already observed in Sect. 4, the comprehension property of our models is crucial to the semantic interpretation of inverse-image relations in such cases. System R of [1] departs from λ2R (and System P) in two main ways. The first is that, in System R, every type A has an associated identity relation A∗ .2 A key rule of System R (written in our notation) is that Θ  x A∗ x, whenever x : A appears anywhere in relational context Θ. This rule breaks the independence between left and right variables in the relational judgements of λ2R. (For example, property (i) of Lemma 7 fails.) The second difference is that System R has an explicit syntax for defining graph relations, rather than the inverse-image construct of λ2R (and System P), which would be more general in that context. Due to the presence of both identity and graph relations, the arguments, in System R, for consequences of parametricity proceed along standard lines [1]. However, System R currently lacks a corresponding semantic story of the kind we have used in this paper in justification of λ2R. In fact, the interplay between models and syntax could be pushed much further than in the present paper. By adding primitive product types to λ2 and λ2R, one can strengthen our full completeness results by obtaining syntactic categories that are initial in an appropriate 2-category of strict structure-preserving morphisms of models. It would be more natural, however, to broaden both the notion of model, by replacing splittings of fibrations with cleavages, and the notion of morphism, by permitting non-strict structure preservation. With such a relaxation, coherence issues arise, but one would expect to obtain (pseudo-) initiality of the syntactic model of λ2R (without any need to extend the syntax with products). For lack of space we have not presented any concrete models in this paper. In fact, any instance of the more elaborate axiomatic structure from [6] can 2

In System P, every type A is itself a relation, which, although called an “identity relation” in [7], has the properties of the relation A in the present paper.

18

N. Ghani et al.

be reconstrued (albeit in a nontrivial way) as a comprehensive λ2 parametricity graph. So our minimal structure at least generalises the known models of parametricity. However, we do not know whether our structure encompasses any genuinely new models of relational parametricity that truly exploit the (potential) added generality of our approach. The results of the present paper should be contrasted with those of other recent work by first two authors and colleagues [9, 10]. In this paper, we have axiomatised category-theoretic structure modelling relational parametricity for the specific type theory λ2, where the resulting structure encompasses both ‘syntactic’ and ‘semantic’ models. In contrast, [9, 10] axiomatise the categorytheoretic structure required on a ‘semantic’ model for Reynolds’ original settheoretic definition of relational parametricity [23] to generalise to the model. Interestingly, the category-theoretic notion of bifibration occurs both as a central ingredient in the axiomatisation of [9, 10], and, in the guise of direct-image relations, as a vital tool in the present paper. A novelty in the present paper is that the bifibrational structure is derived rather than assumed. From a type-theoretic perspective, one advantage of the approach followed in this paper is that the passage from the original type theory (λ2) to the relational version (λ2R) appears not to depend on specific properties of the former, other than that essential use is made of judgemental equality in the formulation of the parametricity rule. We believe that this potential flexibility may be useful for transferring our methods to dependent type theories, where parametricity is an active area of study [2, 4,5, 17]. The proof-relevant setting of dependent type theory, however, requires modifications to our semantic framework. In particular the relational property of parametricity graphs must be relaxed. Ongoing work on a higher-dimensional, proof-relevant form of parametricity may show how to remove this requirement. Acknowledgements. We thank Bob Atkey, Claudio Hermida, Rasmus Møgelberg and the anonymous reviewers for helpful discussions and comments. This research was supported by EPSRC grants GR/A11731/01, EP/E016146/1, EP/K023837/1 and EP/M016951/1.

References 1. Abadi, M., Cardelli, L., Curien, P.-L.: Formal parametric polymorphism. Theor. Comput. Sci. 121(1&2), 9–58 (1993) 2. Atkey, R., Ghani, N., Johann, P.: A relationally parametric model of dependent type theory. In: Jagannathan, S., Sewell, P. (eds.), POPL, pp. 503–515. ACM (2014) 3. Barendregt, H.: Introduction to generalized type systems. J. Funct. Program. 1(2), 125–154 (1991) 4. Bernardy, J.-P., Coquand, T., Moulin, G.: A presheaf model of parametric type theory. In: Ghica, D.R. (ed.), MFPS, ENTCS, pp. 17–33. Elsevier (2015) 5. Bernardy, J.-P., Jansson, P., Paterson, R.: Proofs for free. J. Funct. Program. 22, 107–152 (2012)

Comprehensive Parametric Polymorphism: Categorical Models

19

6. Møgelberg, E.: Categorical models for Abadi and Plotkin’s logic for parametricity. Math. Struct. Comput. Sci. 15, 709–772 (2005) 7. Dunphy, B.: Parametricity as a notion of uniformity in reflexive graphs. Ph.D. thesis, University of Illinois (2002) 8. Dunphy, B., Reddy, U.: Parametric limits. In: LICS, pp. 242–251 (2004) 9. Ghani, N., Johann, P., Nordvall Forsberg, F., Orsanigo, F., Revell, T.: Bifibrational functorial semantics of parametric polymorphism. In: Ghica, D.R. (ed.) MFPS, ENTCS, pp. 67–83. Elsevier (2015) 10. Ghani, N., Nordvall Forsberg, F., Orsanigo, F.: Parametric polymorphism — universally. In: de Paiva, V., de Queiroz, R., Moss, L.S., Leivant, D., de Oliveira, A. (eds.) WoLLIC 2015. LNCS, vol. 9160, pp. 81–92. Springer, Heidelberg (2015) 11. Girard, J.-Y.: Interpr´etation fonctionelle et ´elimination des coupures dans l’arithm´etique d’ordre sup´erieur. Ph.D. thesis, University of Paris VII (1972) 12. Hermida, C.: Fibrational relational polymorphism (2006). http://maggie.cs. queensu.ca/chermida/papers/FibRelPoly.pdf 13. Hermida, C., Reddy, U., Robinson, E.: Logical relations and parametricity – a Reynolds programme for category theory and programming languages. ENTCS 303, 149–180 (2014) 14. Jacobs, B.: Comprehension categories and the semantics of type dependency. Theor. Comput. Sci. 107(2), 169–207 (1993) 15. Jacobs, B.: Categorical Logic and Type Theory. Elsevier, Amsterdam (1999) 16. Kinoshita, Y., O’Hearn, P.W., Power, J., Takeyama, M., Tennent, R.D.: An axiomatic approach to binary logical relations with applications to data refinement. In: Abadi, M., Ito, T. (eds.) Theoretical Aspects of Computer Software. LNCS, vol. 1281, pp. 191–212. Springer, Heidelberg (1997) 17. Krishnaswami, N.R., Dreyer, D.: Internalizing relational parametricity in the extensional calculus of constructions. In: Ronchi, S., Rocca, D. (eds.) CSL, pp. 432–451 (2013) 18. Lawvere, F.W.: Equality in hyperdoctrines and comprehension schema as an adjoint functor. Appl. Categorical Algebra 17, 1–14 (1970) 19. Ma, Q., Reynolds, J.C.: Types, abstraction, and parametric polymorphism, part 2. In: Brookes, S., Main, M., Melton, A., Mislove, M., Schmidt, D. (eds.) Mathematical Foundations of Programming Semantics. LNCS, vol. 598, pp. 1–40. Springer, Heidelberg (1991) 20. O’Hearn, P.W., Tennent, R.D.: Parametricity and local variables. J. ACM 42(3), 658–709 (1995) 21. Plotkin, G., Abadi, M.: A logic for parametric polymorphism. In: Bezem, M., Groote, J.F. (eds.) Typed Lambda Calculi and Applications. LNCS, vol. 664, pp. 361–375. Springer, Heidelberg (1993) 22. Reynolds, J.: Towards a theory of type structure. In: Robinet, B. (ed.) Programming Symposium. LNCS, vol. 19, pp. 408–425. Springer, Heidelberg (1974) 23. Reynolds, J.: Types, abstraction and parametric polymorphism. In: Mason, R.E.A. (ed.), Information Processing, pp. 513–523 (1983) 24. Robinson, E.P., Rosolini, G.: Reflexive graphs and parametric polymorphism. In: LICS, pp. 364–371. IEEE Computer Society (1994) 25. Seely, R.A.G.: Categorical semantics for higher order polymorphic lambda calculus. J. Symbolic Logic 52, 969–989 (1987) 26. Strachey, C.: Fundamental concepts in programming languages. High. Order Symbolic Comput. 13(1–2), 11–49 (2000) 27. Wadler, P.: Theorems for free! In: Stoy, J.E. (ed.) FPCA, pp. 347–359. ACM (1989)

Guarded Dependent Type Theory with Coinductive Types Aleˇs Bizjak1(B) , Hans Bugge Grathwohl1 , Ranald Clouston1 , Rasmus E. Møgelberg2 , and Lars Birkedal1 1

Aarhus University, Aarhus, Denmark {abizjak,hbugge,ranald.clouston,birkedal}@cs.au.dk 2 IT University of Copenhagen, Copenhagen, Denmark [email protected] Abstract. We present guarded dependent type theory, gDTT, an extensional dependent type theory with a ‘later’ modality and clock quantifiers for programming and proving with guarded recursive and coinductive types. The later modality is used to ensure the productivity of recursive definitions in a modular, type based, way. Clock quantifiers are used for controlled elimination of the later modality and for encoding coinductive types using guarded recursive types. Key to the development of gDTT are novel type and term formers involving what we call ‘delayed substitutions’. These generalise the applicative functor rules for the later modality considered in earlier work, and are crucial for programming and proving with dependent types. We show soundness of the type theory with respect to a denotational model.

1

Introduction

Dependent type theory is useful both for programming, and for proving properties of elements of types. Modern implementations of dependent type theories such as Coq [17], Nuprl [11], Agda [21], and Idris [8], have been used successfully in many projects. However, they offer limited support for programming and proving with coinductive types. One of the key challenges is to ensure that functions on coinductive types are well-defined; that is, productive with unique solutions. Syntactic guarded recursion [12], as used for example in Coq [13], ensures productivity by requiring that recursive calls be nested directly under a constructor, but it is well known that such syntactic checks exclude many valid definitions, particularly in the presence of higher-order functions. To address this challenge, a type-based approach to guarded recursion, more flexible than syntactic checks, was first suggested by Nakano [20]. A new modality, written  and called ‘later’ [2], allows us to distinguish between data we have access to now, and data which we will get later. This modality must be used to guard self-reference in type definitions, so for example guarded streams of natural numbers are described by the guarded recursive equation StrgN  N × StrgN asserting that stream heads are available now, but tails only later. c Springer-Verlag Berlin Heidelberg 2016  B. Jacobs and C. L¨ oding (Eds.): FOSSACS 2016, LNCS 9634, pp. 20–35, 2016. DOI: 10.1007/978-3-662-49630-5 2

Guarded Dependent Type Theory with Coinductive Types

21

Types defined via guarded recursion with  are not standard coinductive types, as their denotation is defined via models based on the topos of trees [5]. More pragmatically, the bare addition of  disallows productive but acausal [16] functions such as the ‘every other’ function that returns every second element of a stream. Atkey and McBride proposed clock quantifiers [3] for such functions; these have been extended to dependent types [7, 19], and Møgelberg [19, Theorem2] has shown that they allow the definition of types whose denotation is precisely that of standard coinductive types interpreted in set-based semantics. As such, they allow us to program with real coinductive types, while retaining productivity guarantees. In this paper we introduce the extensional guarded dependent type theory gDTT, which provides a framework where guarded recursion can be used not just for programming with coinductive types but also for coinductive reasoning. As types depend on terms, one of the key challenges in designing gDTT is coping with elements that are only available later, i.e., elements of types of the form A. We do this by generalising the applicative functor structure of  to the dependent setting. Recall the rules for applicative functors [18]: Γ t:A Γ  next t : A

Γ  f : (A → B) Γ  t : A Γ  f  t : B

(1)

The first rule allows us to make later use of data that we have now. The second allows, for example, functions to be applied recursively to the tails of streams. Suppose now that f has type (Πx : A.B), and t has type A. What should the type of f t be? Intuitively, t will eventually reduce to some value next u, and so the resulting type should be (B[u/x]), but if t is an open term we may not be able to perform this reduction. This problem occurs in coinductive reasoning: if, e.g., A is StrgN , and B a property of streams, in our applications f will be a (guarded) coinduction assumption that we will want to apply to the tail of a stream, which has type StrgN . We hence must introduce a new notion, of delayed substitution, similar to let-binding, allowing us to give f  t the type  [x  t] .B binding x in B. Definitional equality rules then allow us to simplify this type when t has form next u, i.e.,  [x  next u] .B ≡ (B[u/x]). This construction generalises to bind a list of variables. Delayed substitution is essential to many examples, as shown in Sect. 3, and surprisingly the applicative functor termformer , so central to the standard presentation of applicative functors, turns out to be definable via delayed substitutions, as shown in Sect. 2. Contributions. The contributions of this paper are: – We introduce the extensional guarded dependent type theory gDTT, and show that it gives a framework for programming and proving with guarded recursive

22

A. Bizjak et al.

and coinductive types. The key novel feature is the generalisation of the ‘later’ type-former and ‘next’ term-former via delayed substitutions; – We prove the soundness of gDTT via a model similar to that used in earlier work on guarded recursive types and clock quantifiers [7, 19]. We focus on the design and soundness of the type theory and restrict attention to an extensional type theory. We postpone a treatment of an intensional version of the theory to future work (see Sects. 7 and 8). In addition to the examples included in this paper, we are pleased to note that a preliminary version of gDTT has already proved crucial for formalizing a logical relations adequacy proof of a semantics for PCF using guarded recursive types by Paviotti et. al. [22]. Note that for space reasons many details appear only in the technical report version of this paper [6].

2

Guarded Dependent Type Theory

gDTT is a type theory with base types unit 1, booleans B, and natural numbers N, along with Π-types, Σ-types, identity types, and universes. For space reasons we omit all definitions that are standard to such a type theory; see e.g. Jacobs [15]. Our universes are a` la Tarski, so we distinguish between types and terms, and have terms that represent types; they are called codes of types and  is the code of the type N. they can be recognised by their circumflex, e.g., N We have a map El sending codes of types to their corresponding type. We follow standard practice and often omit El in examples, except where it is important to avoid confusion.

Fig. 1. Judgements in gDTT.

We fix a countable set of clock variables CV = {κ1 , κ2 , · · · } and a single clock constant κ0 , which will be necessary to define, for example, the function hd in Sect. 5. A clock is either a clock variable or the clock constant; they are intuitively temporal dimensions on which types may depend. A clock context Δ, Δ , · · · is a finite set of clock variables. We use the judgement Δ κ to express that either κ is a clock variable in the set Δ or κ is the clock constant κ0 . All judgements, summarised in Fig. 1, are parametrised by clock contexts. Codes of types inhabit universes UΔ parametrised by clock contexts similarly. The universe UΔ is only well-formed in clock contexts Δ where Δ ⊆ Δ . Intuitively, UΔ contains codes of types that can vary only along dimensions in Δ. We have universe inclusions

Guarded Dependent Type Theory with Coinductive Types

23

from UΔ to UΔ whenever Δ ⊆ Δ ; in the examples we will not write these explicitly. Note that we do not have U Δ : UΔ , i.e., these universes do not form a hierarchy. We could additionally have an orthogonal hierarchy of universes, i.e. 1 2 : UΔ : ···. for each clock context Δ a hierarchy of universes UΔ All judgements are closed under clock weakening and clock substitution. The former means that if, e.g., Γ Δ t : A is derivable then, for any clock variable κ ∈ Δ, the judgement Γ Δ,κ t : A is also derivable. The latter means that if, e.g., Γ Δ,κ t : A is derivable and Δ κ then the judgement Γ [κ /κ] Δ t[κ /κ] : A[κ /κ] is also derivable, where clock substitution [κ /κ] is defined as obvious. The rules for guarded recursion can be found in Figs. 2 and 3; rules for coinductive types are postponed until Sect. 4. Recall the ‘later’ type former , which expresses that something will be available at a later time. In gDTT we have κ  for each clock κ, so we can delay a type along different dimensions. As discussed in the introduction, we generalise the applicative functor structure of κ each  via delayed substitutions, which allow a substitution to be delayed until its substituent is available. We showed in the introduction how a type with a κ single delayed substitution  [x  t] .A should work. However if we have a term κ f with more than one argument, for example of type (Π(x : A).Π(y : B).C), κ t κ u (where κ is the applicative functor and wish to type an application f operation  for clock κ) we may have neither t nor u available now, and so we κ need sequences of delayed substitutions to define the type [x  t, y  u].C. Our concrete examples of Sect. 3 will show that this issue arises in practice. We therefore define sequences of delayed substitutions ξ. The new raw types, terms, and delayed substitutions of gDTT are given by the grammar κ

κ t, u ::= · · · | nextκ ξ.t |  t

A, B ::= · · · |  ξA

ξ ::= · | ξ [x  t] .

κ

Note that we just write A where its delayed substitution is the empty ·, and κ that ξ.A binds the variables substituted for by ξ in A, and similarly for next. The three rules DS-Emp, DS-Cons, and Tf- are used to construct the type κ ξ.A. These rules formulate how to generalise these types to arbitrarily long delayed substitutions. Once the type formation rule is established, the introduction rule Ty-Next is the natural one. κ as With delayed substitutions we can define   κ g f κ t  next .g x. f xt κ Using the rules in Fig. 2 we can derive the following typing judgement for κ

Γ Δ f : ξ.Π(x : A).B κ

κ

Γ Δ t : ξ.A

κ t : ξ[x  t].B Γ Δ f

Ty-

When a term has the form nextκ ξ [x  nextκ ξ.u] .t, then we have enough information to perform the substitution in both the term and its type. The rule

24

A. Bizjak et al.

TmEq-Force applies the substitution by equating the term with the result of an actual substitution, nextκ ξ.t[u/x]. The rule TyEq-Force does the same for its type. Using TmEq-Force we can derive the basic term equality κ κ κ (next ξ.t) ≡ next ξ.(f t). (nextκ ξ.f )

typical of applicative functors [18]. It will often be the case that a delayed substitution is unnecessary, because the variable to be substituted for does not occur free in the type/term. This is what TyEq-  -Weak and TmEq-Next-Weak express, and with these we can justify the simpler typing rule κ

Γ Δ f : ξ.(A → B)

κ

κ

Γ Δ t : ξ.A

κ t : ξ.B Γ Δ f

In other words, delayed substitutions on the type are not necessary when we apply a non-dependent function. Further, we have the applicative functor identity law κ t ≡ t. (nextκ ξ.λx.x)

This follows from the rule TmEq-Next-Var, which allows us to simplify a term nextκ ξ [y  t] .y to t. Sometimes it is necessary to switch the order in the delayed substitution. Two substitutions can switch places, as long as they do not depend on each other; this is what TyEq-  -Exch and TmEq-Next-Exch express. Rule TmEq-Next-Comm is not used in the examples of this paper, but it implies the rule nextκ ξ [x  t] .nextκ x ≡ nextκ t, which is needed in Paviotti’s PhD work. 2.1

Fixed Points and Guarded Recursive Types

In gDTT we have for each clock κ valid in the current clock context a fixed-point combinator fixκ . This differs from a traditional fixed-point combinator in that the type of the recursion variable is not the same as the result type; instead its κ type is guarded with . When we define a term using the fixed-point, we say that it is defined by guarded recursion. When the term is intuitively a proof, we say we are proving by L¨ ob induction [2]. Guarded recursive types are defined as fixed-points of suitably guarded functions on universes. This is the approach of Birkedal and Møgelberg [4], but the generality of the rules of gDTT allows us to define more interesting dependent guarded recursive types, for example the predicates of Sect. 3. We first illustrate the technique by defining the (non-dependent) type of guarded streams. Recall from the introduction that we want the type of guarded κ streams, for clock κ, to satisfy the equation StrκA ≡ A × StrκA . The type A will be equal to El(B) for some code B in some universe UΔ κ where the clock variable κ is not in Δ. We then define the code SA of StrκA in the

Guarded Dependent Type Theory with Coinductive Types

25

Fig. 2. Overview of the new typing rules involving  and delayed substitutions. κ  κ X, where ×  is the code of the (simple) universe UΔ,κ to be SA  fixκ X.B × κ product type. Via the rules of gDTT we can show StrκA  A × StrκA as desired. κ The head and tail operations, hdκ : StrκA → A and tlκ : StrκA → StrκA are simply the first and the second projections. Conversely, we construct streams by pairing. We use the suggestive consκ notation which we define as   κ κ consκ  λ (a : A) as : StrκA . a, as consκ : A → StrκA → StrκA

Defining guarded streams is also done via guarded recursion, for example the stream consisting only of ones is defined as ones  fixκ x.consκ 1 x. The rule TyEq-El- is essential for defining guarded recursive types as fixedpoints on universes, and it can also be used for defining more advanced guarded recursive dependent types such as covectors; see Sect. 3. 2.2

Identity Types

gDTT has standard extensional identity types IdA(t, u) (see, e.g., Jacobs [15]) but with two additional type equivalences necessary for working with guarded dependent types. We write rA t for the reflexivity proof IdA(t, t). The first type equivalence is the rule TyEq-. This rule, which is validated by the model of Sect. 6, may be thought of by analogy to type equivalences often considered in homotopy type theory [24], such as IdA×B( s1 , s2 , t1 , t2 ) ≡ IdA(s1 , t1 ) × IdB(s2 , t2 ).

(2)

There are two important differences. The first is that (2) is (using univalence) a propositional type equality, whereas TyEq- specifices a definitional type equality. This is natural in an extensional type theory. The second difference is that

26

A. Bizjak et al.

Fig. 3. New type and term equalities in gDTT. Rules TyEq--Weak and TmEqNext-Weak require that A and u are well-formed in a context without x. Rules TyEq--Exch and TmEq-Next-Exch assume that exchanging x and y is allowed, i.e., that the type of x does not depend on y and vice versa. Likewise, rule TmEqNext-Comm assumes that exchanging the codomains of ξ and ξ  is allowed and that none of the variables in the codomains of ξ and ξ  appear in the type of u.

there are terms going in both directions in (2), whereas we would have a term κ of type Idκξ.A(nextκ ξ.t, nextκ ξ.u) → ξ.IdA(t, u) without the rule TyEq-. The second novel type equality rule, which involves clock quantification, will be presented in Sect. 4.

3

Examples

In this section we present some example terms typable in gDTT. Our examples will use a term, which we call pη, of type Π(s, t : A × B).IdA(π1 t, π1 s) → IdB(π2 t, π2 s) → IdA×B(t, s). This term is definable in any type theory with a strong (dependent) elimination rule for dependent sums. The second property κ we will use is that StrκA ≡ A × StrκA . Because hdκ and tlκ are simply first and second projections, pη also has type Π (xs, ys : StrκA ) .IdA(hdκ xs, hdκ ys) → IdκStrκ (tlκ xs, tlκ ys) → IdStrκA(xs, ys). A

zipWithκ Preserves Commutativity. In gDTT we define the zipWithκ function which has the type (A → B → C) → StrκA → StrκB → StrκC by κ κ κ tl xs κ tl ys). zipWithκ f  fixκ φ.λxs, ys.consκ (f (hdκ xs) (hdκ ys)) (φ

Guarded Dependent Type Theory with Coinductive Types

27

We show that commutativity of f implies commutativity of zipWithκ f , i.e., that Π(f : A → A → B). (Π(x, y : A).IdB(f x y, f y x)) → Π(xs, ys : StrκA ).IdStrκB(zipWithκ f xs ys, zipWithκ f ys xs) is inhabited. The term that inhabits this type is κ κ κ tl xs κ tl ys). λf.λc.fixκ φ.λxs, ys.pη (c (hdκ xs) (hdκ ys)) (φ κ

Here, φ has type (Π(xs, ys : StrκA ).IdStrκB(zipWithκ f xs ys, zipWithκ f ys xs)) so to type the term above, we crucially need delayed substitutions. An Example with Covectors. The next example is more sophisticated, as it involves programming and proving with a data type that, unlike streams, is dependently typed. Indeed the generalised later, carrying a delayed substitution, is necessary to type even elementary programs. Covectors are the potentially infinite version of vectors (lists with length). To define guarded covectors wefirst need guarded in gDTT is CoNκ   co-natural numbers. The definition κ +  κ X) ; this type satisfies CoNκ ≡ 1 +  CoNκ . Using CoNκ we El fixκ X.(1 κ n), where can define the type family of covectors CoVecκA n  El(CoVec A  κ  κ fixκ φ : (CoNκ → UΔ,κ ) .λ(n : CoNκ ).case n of CoVec A  inl u ⇒ 1 κ

  (φ κ m). inr m ⇒ A × κ . As an example of covecWe will not distinguish between CoVecκA and CoVec A κ tors, we define ones of type Π(n : CoN ). CoVecκN n which produces a covector of any length consisting only of ones: κ m } . ones  fixκ φ.λ(n : CoNκ ).case n of {inl u ⇒ inl ; inr m ⇒ 1, φ

Although this is one of the simplest covector programs one can imagine, it does not type-check without the generalised later with delayed substitutions. The map function on covectors is defined as map : (A → B) → Π(n : CoNκ ). CoVecκA n → CoVecκB n mapf  fixκ φ.λ(n : CoNκ ).case n of inl u ⇒ λ(x : 1).x   κ κ m κ (π2 p) . inr m ⇒ λ p : A × [n  m].(CoVecκA n) . f (π1 p) , φ It preserves composition: the following type is inhabited Π(f : A → B)(g : B → C)(n : CoNκ )(xs : CoVecκA n). IdCoVecκC n(map g n (map f n xs), map (g ◦ f ) n xs)

28

A. Bizjak et al.

by the term λ(f : A → B)(g : B → C).fixκ φ.λ(n : CoNκ ).case n of inl u ⇒ λ(xs : 1).r1 xs κ π2 xs). κ m inr m ⇒ λ(xs : CoVecκA (inr m)).pη (rC g(f (π1 xs))) (φ

4

Coinductive Types

As discussed in the introduction, guarded recursive types on their own disallow productive but acausal function definitions. To capture such functions we need κ to be able to remove . However such eliminations must be controlled to avoid κ κ trivialising . If we had an unrestricted elimination term elim : A → A every κ type would be inhabited via fix , making the type theory inconsistent. κ However, we may eliminate  provided that the term does not depend on the clock κ, i.e., the term is typeable in a context where κ does not appear. Intuitively, such contexts have no temporal properties along the κ dimension, so we may progress the computation without violating guardedness. Figure 4 extends the system of Fig. 2 to allow the removal of clocks in such a setting, by introducing clock quantifiers ∀κ [3,7, 19]. This is a binding construct with associated term constructor Λκ, which also binds κ. The elimination term is clock application. Application of the term t of type ∀κ.A to a clock κ is written as t[κ]. One may think of ∀κ.A as analogous to the type ∀α.A in polymorphic lambda calculus; indeed the basic rules are precisely the same, but we have an additional construct prev κ.t, called ‘previous’, to allow removal of the later κ modality . Typing this new construct prev κ.t is somewhat complicated, as it requires ‘advancing’ a delayed substitution, which turns it into a context morphism (an actual substitution); see Fig. 5 for the definition. The judgement ρ :Δ Γ → Γ  expresses that ρ is a context morphism from context Γ Δ to the context Γ  Δ . We use the notation ρ[t/x] for extending the context morphism by mapping the variable x to the term t. We illustrate this with two concrete examples. First, we can indeed remove later under a clock quantier: κ

force : ∀κ.  A → ∀κ.A

force  λx.prev κ.x[κ] . κ

The type is correct because advancing the empty delayed substitution in  turns it into the identity substitution ι, and Aι ≡ A. The β and η rules (Fig. 6) ensure that force is the inverse to the canonical term λx.Λκ.nextκ x[κ] of type κ ∀κ.A → ∀κ.  A. Second, we may see an example with a non-empty delayed substitution in κ nextκ 0 of type ∀κ.N. Recall that κ is syntactic the term prev κ.nextκ λn.succ n sugar and so more precisely the term is   f  nextκ λn.succ n .f x. (3) prev κ.nextκ x  nextκ 0

Guarded Dependent Type Theory with Coinductive Types Γ

Γ

Δ

Γ

Δ

Γ

Δ,κ

A type

∀κ.A type Γ

Δ

Γ

Δ

Δ ⊆Δ Tf-∀

Δ,κ

t:A

Λκ.t : ∀κ.A Γ Γ

Δ

Δ

Γ Γ

κ

Γ Δ,κ

t:

t : ∀κ.UΔ



∀ t : UΔ

Δ

Ty-Λ Γ

Δ

Δ

Γ Δ

t κ

Δ

29

Ty-∀-code

t : ∀κ.A

: A[κ /κ]

Ty-app

κ

prev κ.t : ∀κ.(A(advκΔ (ξ)))

Ty-prev

Fig. 4. Overview of the new typing rules for coinductive types.

Fig. 5. Advancing a delayed substitution.

Advancing the delayed substitution turns it into the substitution mapping the variable f to the term (prev κ.nextκ λn.succ n)[κ] and the variable x to the term (prev κ.nextκ 0)[κ]. Using the β rule for prev, then the β rule for ∀κ, this simplifies to the substitution mapping f to λn.succ n and x to 0. With this we have that the term (3) is equal to Λκ. ((λn.succ n) 0) which is in turn equal to Λκ.1. An important property of the term prev κ.t is that κ is bound in t; hence prev κ.t has type ∀κ.A instead of just A. This ensures that substitution of terms in types and terms is well-behaved and we do not need the explicit substitutions used, for example, by Clouston et al. [9] where the unary type-former  was used in place of clocks. This binding structure ensures, for instance, that the introduction rule Ty-Λ closed under substitution in Γ . The rule TmEq-∀-fresh states that if t has type ∀κ.A and the clock κ does not appear in the type A, then it does not matter to which clock t is applied, as the resulting term will be the same. In the polymorphic lambda calculus, the corresponding rule for universal quantification over types would be a consequence of relational parametricity. We further have the construct  ∀ and the rule Ty-∀-code which witness that the universes are closed under ∀κ. To summarise, the new raw types and terms, extending those of Sect. 2, are A, B ::= · · · | ∀κ.A

t, u ::= · · · | Λκ.t | t[κ] |  ∀t | prev κ.t

Finally, we have the equality rule TyEq-∀-Id analogous to the rule TyEq. Note that, as in Sect. 2.2, there is a canonical term of type Id∀κ.A(t, s) → ∀κ.IdA(t[κ] , s[κ]) but, without this rule, no term in the reverse direction.

30

4.1

A. Bizjak et al.

Derivable Type Isomorphisms

The encoding of coinductive types using guarded recursive types crucially uses a family of type isomorphisms commuting ∀κ over other type formers [3, 19]. By a type isomorphism A ∼ = B we mean two well-typed terms f and g of types f : A → B and g : B → A such that f (g x) ≡ x and g(f x) ≡ x. The first type isomorphism is ∀κ.A ∼ = A whenever κ is not free in A. The terms g = λx.Λκ.x of type A → ∀κ.A and f = λx.x[κ0 ] of type A → ∀κ.A witness the isomorphism. Note that we used the clock constant κ0 in an essential way. The equality f (g x) ≡ x follows using only the β rule for clock application. The equality g(f x) ≡ x follows using by the rule TmEq-∀-fresh. The following type isomorphisms follow by using β and η laws for the constructs involved. – If κ ∈ A then ∀κ.Π(x : A).B ∼ = Π(x : A).∀κ.B. – ∀κ.Σ (x : A) B ∼ = Σ (y : ∀κ.A) (∀κ.B[y[κ]/x]) . κ – ∀κ.A ∼ = ∀κ.  A. There is an important additional type isomorphism witnessing that ∀κ commutes with binary sums; however unlike the isomorphisms above we require equality reflection to show that the two functions are inverse to each other up to definitional equality. There is a canonical term of type ∀κ.A + ∀κ.B → ∀κ.(A + B) using just ordinary elimination of coproducts. Using the fact that we encode binary coproducts using Σ-types and universes we can define a term com+ of type ∀κ.(A + B) → ∀κ.A + ∀κ.B which is a inverse to the canonical term. In particular com+ satisfies the following two equalities which will be used below. com+ (Λκ.inl t) ≡ inl Λκ.t

5

com+ (Λκ.inr t) ≡ inr Λκ.t.

(4)

Example Programs with Coinductive Types

 in clock context Δ and κ a fresh clock variable. Let A be a type with code A κ Let StrA = ∀κ.StrA . We can define head, tail and cons functions hd : StrA → A tl : StrA → StrA cons : A → StrA → StrA

hd  λxs.hdκ0 (xs[κ0 ]) tl  λxs.prev κ.tlκ (xs[κ]) cons  λx.λxs.Λκ.consκ x (nextκ (xs[κ])).

With these we can define the acausal ‘every other’ function eoκ that removes every second element of the input stream. It is acausal because the second element of the output stream is the third element of the input. Therefore to type the function we need to have the input stream always available, so clock quantification must be used. The function eoκ of type StrA → StrκA is defined as κ κ next ((tl (tl xs)))). eoκ  fixκ φ.λ (xs : StrA ) .consκ (hd xs) (φ

Guarded Dependent Type Theory with Coinductive Types

31

Definitional type equalities: Γ

Δ ⊆Δ

Δ

Γ Γ

Δ

Γ

Δ

Γ

t : UΔ

Δ,κ



El(∀ Λκ.t) ≡ ∀κ. El(t) A type

Δ,κ

Γ

Δ

Γ

Δ

TyEq-∀-el

t : ∀κ.A

Γ

Δ

s : ∀κ.A

∀κ.IdA(t[κ] , s[κ]) ≡ Id∀κ.A(t, s)

TyEq-∀-Id

Definitional term equalities: Γ Γ

Δ Δ

Δ

κ

(Λκ.t) κ

κ∈Δ

Γ

Δ,κ

Δ

A type Γ

Γ

Δ

Δ

Δ,κ κ

Γ Δ

t κ

ξ:Γ

κ

κ∈Δ

TmEq-∀-β

≡ t[κ /κ] : A[κ /κ] Γ

Γ

t:A

Δ

t : ∀κ.A

≡t κ Γ

Γ Δ

Δ

Γ

κ

Δ

Γ

Δ,κ

Δ

κ

:A Γ, Γ

Δ,κ

Γ

Δ,κ

t:

t : ∀κ.A

Λκ.t[κ] ≡ t : ∀κ.A

t:A

prev κ. next ξ.t ≡ Λκ.t(advκΔ (ξ)) : ∀κ.(A(advκΔ (ξ))) Γ

Δ

TmEq-∀-η

TmEq-∀-fresh

TmEq-prev-β

κ

nextκ ((prev κ.t)[κ]) ≡ t :

κ

TmEq-prev-η

Fig. 6. Type and term equalities involving clock quantification.

The result is a guarded stream, but we can easily strengthen it and define eo of type StrA → StrA as eo  λxs.Λκ.eoκ xs. We can also work with covectors (not just guarded covectors as in Sect. 3). This is a dependent coinductive type indexed by conatural numbers which is the type CoN = ∀κ. CoNκ . It is easy to define 0 and succ as 0  Λκ.inl and succ  λn.Λκ.inr (nextκ (n[κ])). Next, we can define a transport function comCoN of type comCoN : CoN → 1 + CoN satisfying comCoN 0 ≡ inl

comCoN (succn) ≡ inr n.

(5)

This function is used to define the type family of covectors as CoVecA n  ∀κ. CoVecκA n where CoVecκA : CoN → UΔ,κ is the term  κ  κ (φ κ (next n)) . 1; inr n ⇒ A× fixκ φ.λ (n : CoN) .case comCoN n of inl ⇒  Using term equalities (4) and (5) we can derive the type isomorphisms CoVecA 0 ≡ ∀κ.1 ∼ =1   κ CoVecA (succn) ≡ ∀κ A ×  (CoVecκA n) ∼ = A × CoVecA n which are the expected properties of the type of covectors. A simple function we can define is the tail function tl : CoVecA (succn) → CoVecA

tl  λv.prev κ.π2 (v[κ]) .

(6)

32

A. Bizjak et al.

Note that (6) is needed to type tl. The map function of type map : (A → B) → Π(n : CoN). CoVecA n → CoVecB n is defined as mapf  λn.λxs.Λκ.mapκ f n (xs[κ]) where mapκ is mapκ : (A → B) → Π(n : CoN). CoVecκA n → CoVecκB n mapκ = λf.fixκ φ.λn.case comCoN n of inl ⇒ λv.v κ κ (next n) κ π2 (v) . inr n ⇒ λv. f (π1 v), φ

5.1

Lifting Guarded Functions

In this section we show how in general we may lift a function on guarded recursive types, such as addition of guarded streams, to a function on coinductive streams. Moreover, we show how to lift proofs of properties, such as the commutativity of addition, from guarded recursive types to coinductive types. Let Γ be a context in clock context Δ and κ a fresh clock. Suppose A and B are types such that Γ Δ,κ A type and Γ, x : A Δ,κ B type. Finally let f be a function of type Γ Δ,κ f : Π(x : A).B. We define L(f ) satisfying the typing judgement Γ Δ L(f ) : Π(y : ∀κ.A).∀κ. (B [y[κ] /x]) as L(f )  λy.Λκ.f (y[κ]). Now assume that f  is another term of type Π(x : A).B (in the same context) and that we have proved Γ Δ,κ p : Π(x : A).IdB(f x, f  x). As above we can give the term L(p) the type Π(y : ∀κ.A).∀κ.IdB[y[κ]/x](f (y[κ]), f  (y[κ])). which by using the type equality TyEq-∀-Id and the η rule for ∀ is equal to the type Π(y : ∀κ.A).Id∀κ.B[y[κ]/x](L(f ) y, L(f  ) y). So we have derived a property of lifted functions L(f ) and L(f  ) from the properties of the guarded versions f and ob induction we prove a property of a f  . This is a standard pattern. Using L¨ function whose result is a “guarded” type and derive the property for the lifted function. For example we can lift the zipWith function from guarded streams to coinductive streams and prove that it preserves commutativity, using the result on guarded streams of Sect. 3.

6

Soundness

gDTT can be shown to be sound with respect to a denotational model interpreting the type theory. The model is a refinement of Bizjak and Møgelberg’s [7] but for reasons of space we leave the description of a full model of gDTT for future work. Instead, to provide some intuition for the semantics of delayed substitutions, we just describe how to interpret the rule

Guarded Dependent Type Theory with Coinductive Types

x : A  B type  t : A  [x  t].B type

33

(7)

in the case where we only have one clock available. The subsystem of gDTT with only one clock can be modelled in the category S, known as the topos of trees [5], the presheaf category over the first infinite ordinal ω. The objects X of S are families of sets X1 , X2 , . . . indexed by the positive integers, together with families of restriction functions riX : Xi+1 → Xi indexed similarly. There is a functor  : S → S which maps an object X to the object !

rX

rX

1 2 X2 ←− X3 ←− · · · 1 ←− X1 ←−

where ! is the unique map into the terminal object. In this model, a closed type A is interpreted as an object of S and the type x : A  B type is interpreted as an indexed family of sets Bi (a), for a in Ai together with maps riB (a) : Bi+1 (a) → Bi (riA (a)). The term t in (7) is interpreted as a morphism t : 1 → A so ti (∗) is an element of Ai (here we write ∗ for the element of 1). The type  [x  t].B type is then interpreted as the object X, defined by X1 = 1

Xi+1 = Bi (ti+1 (∗)).

Notice that the delayed substitution is interpreted by substitution (reindexing) in the model; the change of the index in the model (Bi is reindexed along ti+1 (∗)) corresponds to the delayed substitution in the type theory. Further notice that if B does not depend on x, then the interpretation of   [x  t] .B type reduces to the interpretation B, which is defined to be  applied to the interpretation of B. The above can be generalised to work for general contexts and sequences of delayed substitutions, and one can then validate that the definitional equality rules do indeed hold in this model.

7

Related Work

Birkedal et al. [5] introduced dependent type theory with the  modality, with semantics in the topos of trees. The guardedness requirement was expressed using the syntactic check that every occurrence of a type variable lies beneath a . This requirement was subsequently refined by Birkedal and Møgelberg [4], who showed that guarded recursive types could be constructed via fixed-points of functions on universes. However, the rules considered in these papers do not allow one to apply terms of type (Π(x : A).B), as the applicative functor construction  was defined only for simple function spaces. They are therefore less expressive for both programming (consider the covector ones, and function map, of Sect. 3) and proving, noting the extensive use of delayed substitutions in our example proofs. They further do not consider coinductive types, and so are restricted to causal functions.

34

A. Bizjak et al.

The extension to coinductive types, and hence acausal functions, is due to Atkey and McBride [3], who introduced clock quantifiers into a simply typed setting with guarded recursion. Møgelberg [19] extended this work to dependent types and Bizjak and Møgelberg [7] refined the model further to allow clock synchronisation. Clouston et al. [9] introduced the logic Lgλ to prove properties of terms of the (simply typed) guarded λ-calculus, gλ. This allowed proofs about coinductive types, but not in the integrated fashion supported by dependent type theories. Moreover it relied on types being “total”, a property that in a dependently typed setting would entail a strong elimination rule for , which would lead to inconsistency. Sized types [14] have been combined with copatterns [1] as an alternative type-based approach for modular programming with coinductive types. This work is more mature than ours with respect to implementation and the demonstration of syntactic properties such as normalisation, and so further development of gDTT is essential to enable proper comparison. One advantage of gDTT is that the later modality is useful for examples beyond coinduction, and beyond the utility of sized types, such as the guarded recursive domain equations used to model program logics [23].

8

Conclusion and Future Work

We have described the dependent type theory gDTT. The examples we have detailed show that gDTT provides a setting for programming and proving with guarded recursive and coinductive types. In future work we plan to investigate an intensional version of the type theory and construct a prototype implementation to allow us to experiment with larger examples. Preliminary work has suggested that the path type of cubical type theory [10] interacts better with the new constructs of gDTT than the ordinary Martin-L¨ of identity type. Finally, we are investigating whether the generalisation of applicative functors [18] to apply over dependent function spaces, via delayed substitutions, might also apply to examples quite unconnected to the later modality. Acknowledgements. This research was supported in part by the ModuRes Sapere Aude Advanced Grant and DFF-Research Project 1 Grant no. 4002-00442, both from The Danish Council for Independent Research for the Natural Sciences (FNU). Aleˇs Bizjak was supported in part by a Microsoft Research PhD grant.

References 1. Abel, A., Pientka, B.: Wellfounded recursion with copatterns: A unified approach to termination and productivity. In: ICFP, pp. 185–196 (2013) 2. Appel, A.W., Melli`es, P.A., Richards, C.D., Vouillon, J.: A very modal model of a modern, major, general type system. In: POPL, pp. 109–122 (2007)

Guarded Dependent Type Theory with Coinductive Types

35

3. Atkey, R., McBride, C.: Productive coprogramming with guarded recursion. In: ICFP, pp. 197–208 (2013) 4. Birkedal, L., Møgelberg, R.E.: Intensional type theory with guarded recursive types qua fixed points on universes. In: LICS, pp. 213–222 (2013) 5. Birkedal, L., Møgelberg, R.E., Schwinghammer, J., Støvring, K.: First steps in synthetic guarded domain theory: step-indexing in the topos of trees. LMCS 8(4) (2012) 6. Bizjak, A., Grathwohl, H.B., Clouston, R., Møgelberg, R.E., Birkedal, L.: Guarded dependent type theory with coinductive types (2016). http://arxiv.org/abs/1601. 01586 7. Bizjak, A., Møgelberg, R.E.: A model of guarded recursion with clock synchronisation. In: MFPS (2015) 8. Brady, E.: Idris, a general-purpose dependently typed programming language: design and implementation. J. Funct. Program. 23(5), 552–593 (2013) 9. Clouston, R., Bizjak, A., Grathwohl, H.B., Birkedal, L.: Programming and reasoning with guarded recursion for coinductive types. In: FoSSaCS (2015) 10. Cohen, C., Coquand, T., Huber, S., M¨ ortberg, A.: Cubical type theory: a constructive interpretation of the univalence axiom, unpublished (2015) 11. Constable, R.L., Allen, S.F., Bromley, H.M., Cleaveland, W.R., Cremer, J.F., Harper, R.W., Howe, D.J., Knoblock, T.B., Mendler, N.P., Panangaden, P., Sasaki, J.T., Smith, S.F.: Implementing Mathematics with the Nuprl Proof Development System. Prentice-Hall Inc, Upper Saddle River, NJ, USA (1986) 12. Coquand, T.: Infinite objects in type theory. In: TYPES, pp. 62–78 (1993) 13. Gim´enez, E.: Codifying guarded definitions with recursive schemes. In: TYPES, pp. 39–59 (1995) 14. Hughes, J., Pareto, L., Sabry, A.: Proving the correctness of reactive systems using sized types. In: POPL, pp. 410–423 (1996) 15. Jacobs, B.: Categorical Logic and Type Theory. Studies in Logic and the Foundations of Mathematics, vol. 141. North Holland, Amsterdam (1999) 16. Krishnaswami, N.R., Benton, N.: Ultrametric semantics of reactive programs. In: LICS, pp. 257–266 (2011) 17. The Coq development team: the coq proof assistant reference manual. LogiCal Project (2004). version 8.0. http://coq.inria.fr 18. McBride, C., Paterson, R.: Applicative programming with effects. J. Funct. Program. 18(1), 1–13 (2008) 19. Møgelberg, R.E.: A type theory for productive coprogramming via guarded recursion. In: CSL-LICS (2014) 20. Nakano, H.: A modality for recursion. In: LICS, pp. 255–266 (2000) 21. Norell, U.: Towards a practical programming language based on dependent type theory. Ph.D. thesis, Chalmers University of Technology (2007) 22. Paviotti, M., Møgelberg, R.E., Birkedal, L.: A model of PCF in guarded type theory. In: MFPS (2015) 23. Svendsen, K., Birkedal, L.: Impredicative concurrent abstract predicates. In: Shao, Z. (ed.) ESOP 2014 (ETAPS). LNCS, vol. 8410, pp. 149–168. Springer, Heidelberg (2014) 24. The Univalent Foundations Program: Homotopy Type Theory: Univalent Foundations of Mathematics, Institute for Advanced Study (2013). http:// homotopytypetheory.org/book

Dependent Types and Fibred Computational Effects Danel Ahman1(B) , Neil Ghani2 , and Gordon D. Plotkin1 1

2

LFCS, University of Edinburgh, Edinburgh, Scotland, UK [email protected], [email protected] MSP Group, University of Strathclyde, Glasgow, Scotland, UK [email protected]

Abstract. We study the interplay between dependent types and general computational effects. We define a language with both value types and terms, and computation types and terms, where types depend only on value terms. We use computational Σ-types to account for typedependency in the sequential composition of computations. Our language design is justified by a natural class of categorical models. We account for both algebraic and non-algebraic effects. We also show how to extend the language with general recursion, using continuous families of cpos.

1

Introduction

While dependent types have proven very useful on their own, for both programming and theorem proving, one also seeks a general way to combine them with computational effects, such as I/O, state, continuations, or recursion, so as to write more practical, concise, or clearer programs. However, despite the study of each of the two fields being well advanced, their combination presents difficulties, as recognised already by Moggi [25]. One puzzling problem is what to make of a type A(M ) if M can raise an effect. We do not know a general denotational semantics for such types, though there may be one (there is one for local names [28]). Pragmatically, the situation depends on the nature of the computational effects considered. For effects not requiring interaction with the program runtime, such as local names [28] or general recursion [7], one need not restrict M , as computing A(M ) then only depends on static information. However, this does not work well in general, as some effects, e.g., I/O, do crucially depend on interaction with the program runtime, and so then will the computation of A(M ). As one consequence, extending a coarse-grained language, e.g., Moggi’s computational lambda calculus [24], with dependent types seems not to give a general solution, as the natural elimination rule for Π-types produces types of the form A(M ) (and cf. Levy [20, Sect. 12.4.1]). D. Ahman—This work was funded by a University of Edinburgh PhD scholarship and by the scholarship program Kristjan Jaak, which is funded and managed by Archimedes Foundation in collaboration with the Estonian Ministry of Education and Research. c Springer-Verlag Berlin Heidelberg 2016  B. Jacobs and C. L¨ oding (Eds.): FOSSACS 2016, LNCS 9634, pp. 36–54, 2016. DOI: 10.1007/978-3-662-49630-5 3

Dependent Types and Fibred Computational Effects

37

A natural move to try to solve these problems is to allow types to depend only on value terms. This is the route we take in this paper and it does lead to a natural general semantics. While at first such a choice might seem limiting, we recover dependency on the statically available information about computational effects by inspecting the structure of thunked computations. To ensure that types depend only on values, we make a clear distinction between values and computations, using separate classes of value types A and computation types C, as in Call-By-Push-Value (CBPV) [20] and the Enriched Effect Calculus (EEC) [9]. Then the variables in types range only over value types, as desired. However, a further problem then arises: the usual typing rule Γ  M : F A Γ, x : A  N (x) : C(x) (∗) Γ  M to x in N (x) : C(x) for sequential composition of computations is not correct as x may occur freely in C(x) in the conclusion. An evident move here is to prohibit x from occurring freely in C, as, e.g., advocated by Levy [20, Sect. 12.4.1] and Brady [5]. However, this seems too restrictive: to make the most of dependent types, it is desirable for the type of an effectful program to depend on values produced by preceding computations. As an example, consider combining monadic parsing [16] with dependent types and applying it to the parsing of well-typed syntax. Then it is natural to decompose the parsing of function applications into a parser for the function and a parser for the argument, where the type of the latter crucially depends on the domain of the type of the parsed function. The approach we take is to keep the restricted form of (*), but to also introduce computational Σ-types Σx : A.C(x), whose use is inspired by the algebraic treatment of computational effects [31]. To explain these, suppose computation types denote algebras for the algebraic theory of boolean-valued read-only store [35, Sect. III.A]. Now let us consider the composite effectful program   (return 2) ? (return 3) to x in M (x) where x : Nat  M (x) : C(x). Here we see that, after looking up the bit with ?, this program either evaluates as M (2), which has type C(2), or as M (3), which has type C(3). So the whole computation yields an element of the coproduct of algebras C(2) + C(3). This pattern also recurs with other computational effects, and dependency on value types other than natural numbers: hence the computational Σ-types. These types enable us to type sequential composition by “closing-off” the free variable x in C(x): using the hypothesis of (*) we derive Γ, x : A  x, N (x) : Σx : A.C(x) using the introduction rule for computational Σ-types, and then derive Γ  M to x in x, N (x) : Σx : A.C(x) using the restricted form of (*).

38

D. Ahman et al.

We aim to put these ideas together at a foundational level, comparable with the levels at which the two fields have been studied separately. In contrast, the existing work on combining dependent types with first-class computational effects has concerned only particular kinds of effects (e.g., [7, 26, 28, 37]). Other authors have used dependent types in existing languages to represent effects using DSLs (e.g., [5,13, 23]). There has, however, been no work combining dependent types with general, first class computational effects. In Sect. 2, we define a small dependently-typed language with computational effects, combining features from Martin-L¨ of type theory (MLTT) [22] and computational languages separating values and computations, such as CBPV or EEC. The language design we propose is justified in Sect. 3 in terms of a sound and complete interpretation in a class of categorical models naturally combining (i) comprehension categories arising from the semantics of dependent types; and (ii) adjunctions arising from the semantics of computational effects. In Sect. 4, we extend the language and its semantics with algebraic effects. In Sect. 5, we extend the language and its semantics with general recursion.

2

A Dependently-typed Effectful Language

We now define the syntax and equational theory of our dependently-typed effectful language, making a clear distinction between values and computations, at both type and term levels: the value fragment of our language is based on MLTT; the computational fragment is greatly influenced by CBPV and EEC. Variables. We assume two countable sets: of value variables, ranged over by x, y, . . .; and of computation variables, ranged over by z, . . .. The former are treated as in MLTT: they are intuitionistic, and enjoy structural rules of weakening and contraction. The latter, on the other hand, are treated linearly, as in EEC, and play a crucial role in the elimination rule for computational Σ-types. Types. Our value types A, B, . . . are given as in MLTT, except for the type U C of thunks, familiar from CBPV. To keep the presentation simple and focussed on the computational fragment of our language, we omit value sum types and general inductive types, both of which can be easily added. Our computation types C, D, . . . generalise those of CBPV and EEC. The grammar of types is A ::= Nat | 1 | Σx : A.B | Πx : A.B | IdA (V, W ) | U C

C ::= F A | Σx : A.C | Πx : A.C

Here, F A is the type of computations returning values of type A. The computational Σ- and Π-types are the natural dependently-typed generalisations of EEC’s computational tensor and function types: !A ⊗ C, A → C. Rules defining well-formed value contexts  Γ and types Γ  A, Γ  C are given in Fig. 1. Terms. We let V, W, . . . to range over value terms. These are given as in MLTT, except for thunked computations thunk M , familiar from CBPV. Compared to CBPV, we include both value functions and complex values to accommodate pure programs on which the types of our language could depend. Regarding effectful programs, we further distinguish between computation terms, ranged

Dependent Types and Fibred Computational Effects

39

over by M, N, . . . and homomorphism terms, ranged over by K, L, . . .. We decorate our terms with certain type-annotations, so that we can later define the denotational semantics on raw syntax — a standard approach in the literature [15, 36]. We omit these annotations in informal discussion. The grammar of terms is V ::= x | zero | succ V | recx.A (Vz , y1 .y2 .Vs , W ) |  | λx : A.V | V (W )(x:A).B | V, W (x:A).B | fst (x:A).B V | snd (x:A).B V | thunk M | refl V | indA (x1 .x2 .x3 .B, y.W, V1 , V2 , Vp ) M ::= return V | M to x : A inC N | force C V | λx : A.M | M (V )(x:A).C | V, M (x:A).C | M to x : A, z : C inD K K ::= z | K to x : A inC M | λx : A.K | K(V )(x:A).C | V, K(x:A).C | K to x : A, z : C inD L

Well-formed value types Γ, x : A Γ Γ Γ

Nat

Γ

1

Γ

Well-formed value contexts Γ x ∈ Vars(Γ ) Γ Γ, x : A

B

Σx : A.B A

Γ, x : A Γ

B

Γ

Πx : A.B

V :A Γ W :A Γ IdA (V, W )

Γ Γ

Well-formed computation types Γ, x : A C Γ, x : A A Γ Γ

FA

Γ

Σx : A.C

Γ

C UC C

Πx : A.C

Fig. 1. Well-formed contexts and types.

Our computation terms share many similarities with those in CBPV, but additionally include the introduction and elimination rules for computational Σ-types, given by pairing V, M  and pattern-matching M to x, z in K, whose syntax is similar to that for computational tensor types !A ⊗ C in EEC. As with the linear terms in EEC, our homomorphism terms contain computation variables z used linearly. From an operational perspective, the typing rules for homomorphism terms ensure that a computation bound to z always happens first in a well-typed term containing it. So, when eliminating a pair V, M , M always happens before the rest of K in the compound term V, M  to x, z in K, thus preserving the intended left-to-right evaluation order. This use of computation variables ensures that homomorphism terms denote algebra homomorphisms in the examples based on Eilenberg-Moore algebras of monads, or on algebraic effects [31]: hence the name. Similar forms of linearity are also present in CBPV with stacks [20, Sect. 2.3.4]; indeed, homomorphism terms can be viewed as a programmer-friendly syntax for dependently-typed stack terms. Well-typed value terms Γ  V : A, well-typed computation terms Γ  M : C, and well-typed homomorphism terms Γ | z : C  K : D are defined in Fig. 2. The linear use of computation variables is also reminiscent of recent work on combining dependent and linear types in languages with distinguished intuitionistic and linear fragments [19, 38]. That work is designed to capture the

40

D. Ahman et al.

Type conversion rules Γ

Γ

V :A Γ A=B Γ V :B

M :C Γ

Γ

Γ | z :C

C=D

M :D

K : D1

Γ

Γ | z :C

D1 = D2

K : D2

Well-typed value terms Γ

Γ x:A ∈ Γ Γ x:A A

Γ, x : Nat

Γ

Γ

M :C

Vz : A[zero/x] Γ

Γ

V :A

Γ, x : A

Γ

V, W

B

Γ

(x:A).B

Γ, x : A Γ

Γ

thunk M : U C

Γ, y1 : Nat, y2 : A[y1 /x]

W : B[V /x]

Γ

: Σx : A.B

Γ

V :B

return V : F A

Γ

M : Πx : A.C Γ

Γ

M : Σx : A.C Γ

Γ | z :C

Γ

C Γ

Γ, x : A V :A

N :C

Γ, x : A

Γ

D

Γ

Γ, x : A

Γ

V, M

Γ, x : A | z : C

K : Σx : A.D 1

Γ | z1 : C

Γ, x : A

D

V, K

K : FA

Γ | z :C

Γ V :A refl V : IdA (V, V )

Γ

C

K:D

Γ

Γ | z :C

(x:A).D

D

K : D[V /x]

Γ, x : A

M : C[V /x] V : UC

force C V : C

Γ, x : A | z2 : D 1

L : D2

K to x : A, z2 : D 1 inD2 L : D 2

: Σx : A.D

K to x : A inD M : D

D2

M :C

λx : A.M : Πx : A.C

: Σx : A.C Γ

Γ Γ

Γ

(x:A).C

M to x : A, z : C inD K : D

Γ | z :C Γ | z :C

W : Nat

B Γ, y : A W : B[y/x1 , y/x2 , refl y/x3 ] V2 : A Γ Vp : IdA (V1 , V2 )

V :A

z:C

V :A

Γ

Γ V : Σx : A.B snd(x:A).B V : B[fst(x:A).B V /x]

Γ

M to x : A inC N : C

Well-typed homomorphism terms Γ | z1 : C Γ C Γ

V : Σx : A.B fst(x:A).B V : A

M (V )(x:A).C : C[V /x]

Γ

V : Nat succ V : Nat

indA (x1 .x2 .x3 .B, y.W, V1 , V2 , Vp ) : B[V1 /x1 , V2 /x2 , Vp /x3 ]

Well-typed computation terms Γ M : FA V :A Γ Γ

Γ Γ

Vs : A[succ y1 /x]

Γ V : Πx : A.B Γ W :A Γ V (W )(x:A).B : B[W/x]

Γ, x1 : A, x2 : A, x3 : IdA (x1 , x2 ) V1 : A Γ Γ

Γ

Γ zero : Nat

Γ

:1

recx.A (Vz , y1 .y2 .Vs , W ) : A[W/x]

λx : A.V : Πx : A.B

Γ

Γ

M :D

Γ

C

Γ, x : A | z : C

Γ | z :C Γ | z :C Γ | z :C

K:D

λx : A.K : Πx : A.D

K : Πx : A.D

Γ

V :A

K(V )(x:A).D : D[V /x]

Fig. 2. Well-typed terms.

adjunction models of intuitionistic linear logic [3]; in contrast, our language is designed to capture the computational nature of certain adjunctions. Equations. We equip our language with an equational theory, consisting of equations between well-formed types, written Γ  A = B and Γ  C = D; and welltyped terms, written Γ  V = W : A, Γ  M = N : C and Γ | z : C  K = L : D. The equations between types consist of reflexivity equations for Nat and 1, and congruence rules for all the other type formers. We omit the equations between value terms as they are standard from MLTT with natural numbers and intensional identity types [15, 22]. The rules for equations between computation and homomorphism terms are given in Fig. 3. We leave the well-typedness assumptions about the constituent terms implicit, and omit type conversion, equivalence, and congruence rules. Many of the equations in Fig. 3 are familiar from EEC, modulo the dependent-typing. Compared to other computational languages, such as [24], standard equations that may seem missing from the theory, such as associativity of sequential composition, are in fact derivable.

Dependent Types and Fibred Computational Effects

41

Equations involving thunking and forcing Γ

thunk (force C V ) = V : U C

Γ

force C (thunk M ) = M : C

Equations between well-typed computation terms Γ Γ

return V to x : A inC M = M [V /x] : C

Γ

V, M to x : A, z : C inD K = K[V /x, M/z] : D

Γ

M to x : A inC K[return x/z] = K[M/z] : C Γ

M to x : A, z2 : C inD K[ x, z2 /z1 ] = K[M/z1 ] : D

(λx : A.M )(V )(x:A).C = M [V /x] : C[V /x] Γ

M = λx : A.M (x)(x:A).C : Πx : A.C

Equations between well-typed homomorphism terms Γ | z1 : C Γ | z1 : C Γ | z1 : C

K to x : A inD L[return x/z2 ] = L[K/z2 ] : D

V, K to x : A, z2 : D 1 inD2 L = L[V /x, K/z2 ] : D 2

K to x : A, z3 : D 1 inD2 L[ x, z3 /z2 ] = L[K/z2 ] : D 2

Γ | z1 : C

(λx : A.K)(V )(x:A).D = K[V /x] : D[V /x]

Γ | z1 : C

K = λx : A.K(x)(x:A).D : Πx : A.D

Fig. 3. Fragment of the equational theory.

Some meta-theory. The substitution of value terms for value variables has a straightforward mutually recursive definition. We write A[V /x] for the substitution of V for x in A. The substitution of computation and homomorphism terms for computation variables is also routine, recursing only in the sub-terms where linearly used computation variables can appear. We write K[M/z] for the substitution of M for z in K. Then, standard weakening and substitution rules for value variables are admissible for all judgments of our language. In addition, further substitution rules for computation variables are admissible for judgments Γ | z : C  K : D and Γ | z : C  K = L : D, covering the substitution of both computation and homomorphism terms for computation variables.

3

Denotational Semantics

The denotational semantics of our language is based on standard fibred category theory. To make our work more accessible, we recall some preliminaries of this theory and suggest [18] for more details. Fibred category theory provides a natural framework for developing the semantics of dependently-typed languages, where: (i) functors model type-dependency; (ii) split fibrations model substitution; and (iii) closed comprehension categories model Σ- and Π-types. The ideas we develop can also be expressed in terms of other models of dependent types, such as categories with families, or categories with attributes [15, 27]. 3.1

Fibred Category Theory Preliminaries

Fibrations. Given a functor p : E −→ B, a morphism g : A −→ B is called a Cartesian lifting of f : X −→ Y if p(g) = f and for all i : C −→ B and j : p(C) −→ X, such that p(i) = f ◦ j in B, there exists a unique h : C −→ A over j such that g ◦ h = i. The functor p : E −→ B is called a fibration if for every B in E and f : X −→ p(B) in B there exists a Cartesian lifting g : A −→ B of f in E. A morphism f : A −→ B in E is called vertical if p(A) = p(B) = X and p(f ) = idX . For any X in B, we write EX for the fibre over X, i.e., for

42

D. Ahman et al.

the subcategory of E consisting of objects over X and vertical morphisms. A fibration is called cloven if it comes with a choice of Cartesian liftings. We write f (B) : f ∗ (B) −→ B for the chosen Cartesian lifting of f : X −→ p(B). In cloven fibrations, every B-morphism f : X −→ Y determines a reindexing functor f ∗ : EY −→ EX , satisfying (idX )∗ ∼ = idEX and (g ◦ f )∗ ∼ = f ∗ ◦ g ∗ . A cloven fibration is said to be split if these two isomorphisms are identities. Given split fibrations p : V −→ B and q : C −→ B, a split fibred functor F : p −→ q is given by a functor F : V −→ C, such that q ◦F = p and F preserves the chosen Cartesian morphisms on-the-nose. Given two split fibred functors F, G : p −→ q, a split fibred natural transformation α : F ⇒ G is given by a natural transformation α : F ⇒ G, in which every component of α is vertical. A split fibred adjunction F U : q −→ p is given by split fibred functors F : p −→ q and U : q −→ p, together with split fibred natural transformations η : idV −→ U ◦ F and ε : F ◦ U −→ idC , subject to the standard unit-counit laws for adjunctions. Comprehension categories. A (split) comprehension category with unit is given by a (split) fibration p : E −→ B, together with a comprehension-admitting terminal object functor 1 : B −→ E, i.e., 1 has a right adjoint {−} : E −→ B; it is said to π(−) 1{−} ) : E −→ B→ is full and faithful. For be full when the functor A → p(εA all A in E, the B-morphism πA : {A} −→ p(A) is called a projection map. The ∗ : Ep(A) −→ E{A} is called the weakening corresponding reindexing functor πA functor. For every comprehension category with unit p : E −→ B, we have an isomorphism Ep(A) (1p(A) , A) ∼ = {g : p(A) −→ {A} | πA ◦ g = idp(A) }, for all A in E. As a notational convention, we write s(f ) : p(A) −→ {A} for the section 1{−} . corresponding to the global element f : 1p(A) −→ A, given by {f } ◦ ηA A comprehension category with unit p : E −→ B is said to have dependent ∗ have right products (resp. weak dependent sums) when the weakening functors πA adjoints ΠA (resp. left adjoints ΣA ), for all A in E, satisfying the Beck-Chevalley condition: for all Cartesian morphisms f : A −→ B, the canonical natural transformation (p(f ))∗ ◦ ΠB −→ ΠA ◦ {f }∗ (resp. ΣA ◦ {f }∗ −→ (p(f ))∗ ◦ ΣB ) is an isomorphism. A comprehension category with unit p : E −→ B is said to have strong dependent sums when it has weak dependent sums, s.t. for all B in E{A} , ∗ Σ  πA

the morphism {πA (ΣA B) ◦ ηB A

} : {B} −→ {ΣA B} is an isomorphism.

Split closed comprehension categories. In order to define fibred adjunction models in Sect. 3.2, we use a particularly well-behaved class of comprehension categories (from the perspective of interpreting type theory), namely, those that are split and closed. A split closed comprehension category (SCCompC) is a split full comprehension category with unit p : E −→ B, where the base category B has a terminal object; the fibred terminal objects are preserved on-the-nose by reindexing; and which has dependent products and strong dependent sums, for which the isomorphisms in the Beck-Chevalley conditions are identities. Natural numbers. A SCCompC p : E −→ B is said to support weak natural numbers if there exists an object N in E1 and vertical morphisms zero : 11 −→ N, succ : N −→ N, s.t. for all X in B, A in E{!∗X (N)} , hz : 1X −→ (s(!∗X (zero)))∗ (A) in ∗ ({!∗X (succ)}∗ (A)) in E{A} , there exists h : 1{!∗X (N)} −→ A EX and hs : 1{A} −→ πA ∗ ({!∗X (succ)}∗ (h)) = hs . in E{!∗X (N)} , satisfying (s(!∗X (zero)))∗ (h) = hz and πA

Dependent Types and Fibred Computational Effects

43

We present N axiomatically rather than using weak initial algebras since our language and its models do not assume coproducts. Moreover, discussing the semantics of inductive types and their fibred induction principles in full generality [10] would digress too much from our central theme. Identity types. Following the axiomatic presentation given by Warren [39], a SCCompC p : E −→ B is said to support identity types, if, for all A in E, there ∗ (IdA ) in E{A} , such that for exists an object IdA in E{πA∗ (A)} , and rA : 1{A} −→ δA ∗ all B in E{IdA } and f : 1{A} −→ (s(rA )) ({δA (IdA )}∗ (B)) in E{A} , there exists iA,B (f ) : 1{IdA } −→ B in E{IdA } , satisfying (s(rA ))∗ ({δA (IdA )}∗ (iA,B (f ))) = f . These identity types are also required to satisfy a split Beck-Chevalley condition: for all Cartesian morphisms f : A −→ B, we must have {f  }∗ (IdB ) = IdA in ∗ ∗ (A) −→ πB (B) is the unique mediating morphism over E{πA∗ (A)} , where f  : πA ∗ {f }, arising from πB (B) : πB (B) −→ B being a Cartesian morphism. As in [18, Sect. 9.3.5], the diagonal morphisms δA arise from pullback squares of the form

3.2

Interpretation of Our Language in Fibred Adjunction Models

A fibred adjunction model is given by a SCCompC p : V −→ B, a split fibration q : C −→ B, and a split fibred adjunction F U : q −→ p, such that p supports identity types and weak natural numbers (in the sense of Sect. 3.1), and q supports split dependent products and sums with respect to p as depicted in

The split dependent products and sums in q, with respect to p, are defined as ∗ : Cp(A) −→ C{A} , required left and right adjoints to the weakening functors πA to satisfy the analogues of the split Beck-Chevalley conditions from Sect. 3.1. Given a SCCompC p : V −→ B that supports identity types and weak natural numbers, we can always pick the identity adjunction idV idV : V −→ V to construct a corresponding “effect-free” fibred adjunction model. Further, we can construct a restricted form of adjunction models (without identity types) from models of EEC with weak natural numbers [9], i.e., from D-enriched adjunctions FEEC UEEC : E −→ D, where D is Cartesian closed and has a weak NNO, E is D-enriched and has all D-tensors and -cotensors. These models are based on a computational variant q : s(D, E) −→ D of the the simple fibration

44

D. Ahman et al.

p : s(D) −→ D [18, Theorem 10.5.5]. In particular, the objects of s(D, E) are pairs (X, C) of a D-object X and a E-object C; and the morphisms (X, C) −→ (Y, D) are pairs (f, g) of morphisms, with f : X −→ Y in D and g : X ⊗ C −→ D in E. Interpretation. Following Streicher [36] and Hoffmann [15], we define the interpretation of our language in fibred adjunction models by first giving a partial interpretation function − on raw syntax and then proving that − is defined on well-formed expressions. We do so because of the well-known issue in interpreting dependently-typed languages: as the derivations of judgments are not unique, due to the type conversion rules, defining the interpretation on the derivations would require simultaneously proving that the definition is coherent. If defined, − maps a context Γ to an object Γ  in B; a pair of a context Γ and a value type A to an object Γ ; A in VΓ  ; a pair of a context Γ and computation type C to an object Γ ; C in CΓ  ; a pair of a context Γ and a value term V to an object A and a morphism Γ ; V  : 1Γ  −→ A in VΓ  ; a pair of a context Γ and a computation term M to an object C in CΓ  and a morphism Γ ; M  : 1Γ  −→ U (C) in VΓ  ; and a triple of a context Γ , a computation type C and a homomorphism term K to an object D and a morphism Γ ; C; K : Γ ; C −→ D in CΓ  . We define these different cases of − simultaneously by induction on the depth of the argument expressions. In the definition, it is convenient to use Kleene equality where two sides of an equation are equal if either they are both defined and equal or they are both undefined. For contexts, − is defined as · 1

Γ, x : A {Γ ; A}

if x ∈ Vars(Γ )

For types, we illustrate the definition of − with the following example cases: Γ ; IdA (V, W ) (s(Γ ; V ), s(Γ ; W ))∗ (IdΓ ;A ) Γ ; Nat !∗Γ  (N) Γ ; Σx : A.C ΣΓ ;A (Γ, x : A; C) Γ ; Πx : A.C ΠΓ ;A (Γ, x : A; C) ∗ where the pairing morphism s(Γ ; V ), s(Γ ; W ) : Γ  −→ {πΓ ;A (Γ ; A)} is the unique mediating morphism into the pullback square from Sect. 3.1, for s(Γ ; V ) and s(Γ ; W ). We also note that Γ ; IdA (V, W ) is conditional on the codomains of the morphisms given by Γ ; V  and Γ ; W  being equal to Γ ; A. For terms, we present the definition only for homomorphism terms, given by idΓ ;C

Γ ; C; z  Γ ; C −−−−→ Γ ; C

Γ,x:A;M †

Γ ;C;K

Γ ; C; K to x : A inD M   Γ ; C −−−−→ F (Γ ; A) −−−−− −−→ Γ ; D Γ ;C;K



∗ s(Γ ;V )∗ (η Σπ ) Γ,x:A;D

−−→ Γ ; C; V, K(x:A).D   Γ ; C −−−−→ s(Γ ; V ) (Γ, x : A; D) −−−−−−−−−−−−−−− =

∗ s(Γ ; V )∗ (πΓ ;A (ΣΓ ;A (Γ, x : A; D))) −→ ΣΓ ;A (Γ, x : A; D)  

Γ ; C; K to x : A, z : D 1  inD

Γ,x:A;D 1 ;L

Γ ;C;K

2

L  Γ ; C −−− −−→ ΣΓ ;A (Γ, x : A; D 1 ) −−−−−−−−−−→ Γ ; D 2 

λΓ,x:A;C;K

−−→ ΠΓ ;A D Γ ; C; λx : A.K  Γ ; C −−−−−−− =

∗ Γ ; C; K(V )(x:A).D   Γ ; C −→ s(Γ ; V )∗ (πΓ ;A (Γ ; C))

s(Γ ; V )



∗ (πΓ ;A (ΠΓ ;A (Γ, x : A; D)))

s(Γ ;V )∗ (π ∗ (Γ ;C;K)) Γ ;A

−−−−−−−−−−−−−−−−−−− −−→

∗ s(Γ ;V )∗ (επ Π ) Γ,x:A;D

−−−−−−−−−−−−−−−−→ s(Γ ; V )∗ (Γ, x : A; D)

Dependent Types and Fibred Computational Effects

45

  where the vertical morphisms Γ, x : A; M † , Γ, x : A, D1 ; L and λΓ ; C; K are derived using the adjunctions involved in the definition of fibred adjunction models. Similarly to [15, 36], we make some of the cases of the definition (implicitly) conditional on the denotations of sub-terms having a particular form, based on the type-annotations with which we have decorated our terms. Theorem 1 (Soundness). The partial interpretation function − is defined on well-formed contexts, well-formed types and well-typed terms. In addition, the interpretation identifies types and terms that are equal in the equational theory. Similarly to [15, 36], the proof of soundness relies on lemmas relating weakening and substitution in the syntax to reindexing in fibred adjunction models. The classifying model. We now show that the interpretation of our language in fibred adjunction models is complete by constructing its classifying model. First, in the classifying fibred adjunction model the objects of B are given by equivalence classes of well-formed value contexts Γ . The morphisms Γ1 −→ Γ2 are given by equivalence classes of tuples of value terms V = (V1 , . . . , Vm ), where Γ2 = y1 : B1 , . . . , ym : Bm and Γ1  Vi : Bi [V1 /y1 , . . . , Vi−1 /yi−1 ], for all 1 ≤ i ≤ m. Next, the objects of the total category V are given by equivalence classes of value types Γ  A and its morphisms Γ1  A −→ Γ2  B by equivalence classes of tuples of value terms (V , V ), where V are typed as in B, and V is typed as Γ1 , x : A  V : B[V /Γ2 ]. The objects and morphisms of C are defined similarly: as equivalence classes of computation types Γ  C; and as equivalence classes of tuples of terms (V , K), where K is typed as Γ1 | z : C  K : D[V /Γ2 ]. The fibrations p and q are defined by context projections, i.e., by p(Γ  A) = Γ . The various adjunctions involved in the definition of fibred adjunction models are defined in terms of their syntactic counterparts. For example, the split fibred adjunction F U : q −→ p is defined using the types F A and U C, given by F (Γ  A) = Γ  F A U (Γ  C) = Γ  U C

F (V , V ) = (V , z to y : A inFB[V/Γ2 ] return V [y/x]) U (V , K) = (V, thunk K[force C x/z])

The identity types and natural numbers are also given in terms of their syntactic counterparts: the object IdΓ A in V{πΓ∗ A (Γ A)} is Γ, x : A, y : A  IdA (x, y). Proposition 1. The above definitions, based on the syntax of our language, constitute a fibred adjunction model, called the classifying model of our language. Finally, we can use this result to prove the completeness of the interpretation. Theorem 2 (Completeness). If two types or two terms of our language are identified in all fibred adjunction models, they are equal in the equational theory.

46

3.3

D. Ahman et al.

Fibred Adjunction Models Based on the Families Fibration

We now discuss some examples of fibred adjunction models based on the prototypical model of dependent types, the families fibration p : Fam(D) −→ Set. The objects of Fam(D) are given by pairs (X, A) of a set X and an X-indexed family of D-objects A : X −→ ob(D); the morphisms (X, A) −→ (Y, B) are pairs (f, {gx }x∈X ) of a function f : X −→ Y and a X-indexed family of D-morphisms {gx : A(x) −→ B(f (x))}x∈X . The functor p is defined by first projection, i.e. by p(X, A) = X. In fact, p is a split fibration: the reindexing functors f ∗ are defined by pre-composition, i.e. by f ∗ (Y, B) = (X, B ◦ f ) for all f : X −→ Y . In our examples, we take D to be Set. In this case, p is a SCCompC [18, Sect. 10.5]. The examples we discuss below are instances of the following general result, building on the fact that adjunctions can be lifted to families fibrations [18, Example 1.8.7(i)]. Theorem 3. Given F U : E −→ Set, such that E has both set-indexed products and coproducts, the fibrations p : Fam(Set) −→ Set and q : Fam(E) −→ Set, together with the pointwise lifting of F U , determine a fibred adjunction model. In Theorem 3, the set-indexed products and coproducts in E are assumed to exist in order to define the dependent products and dependent sums in q as   Π(X,A) ( x∈X A(x), C) = (X, x → a∈A(x) C(x, a))   Σ(X,A) ( x∈X A(x), C) = (X, x → a∈A(x) C(x, a)) ˆ of F U is defined by composition, i.e., Fˆ (X, A) = (X, F ◦A). The lifting Fˆ U The first collection of examples we discuss are based on Eilenberg-Moore algebras (EM-algebras) for a monad (T, η, μ) on Set. As standard, we write SetT for the category of EM-algebras. Its objects are given by pairs (X, α) of a set X and a function α : T X −→ X such that α ◦ T (α) = α ◦ μX and α ◦ ηX = idX ; its morphisms (X, α) −→ (Y, β) are given by functions f : X −→ Y such that β ◦ T (f ) = f ◦ α. There is a canonical EM-adjunction F T U T : SetT −→ Set. We note that SetT has both set-indexed products and coproducts: SetT is complete [4, Theorem 4.3.1] and, assuming the axiom of choice (AC), SetT is also cocomplete [4, Theorem 4.3.5]. We note that, without assuming AC, it would suffice to assume that SetT has reflexive coequalizers for it to be cocomplete, as T first observed by Linton [21]. The set-indexed products  in Set are defined from  the set-indexed products in Set by i∈I (Xi , αi ) = ( i∈I Xi , αi ◦ T (proji )i∈I ). As shown by Linton [21], the set-indexed coproduct   i∈I (Xi , αi ) can be defined as the reflexive coequalizer e : F T ( i∈I Xi ) −→ i∈I (Xi , αi ) of the diagram 

FT(

i∈I Xi )

F T ([inji ◦αi ]i∈I )

F T ([inji ◦ηXi ]i∈I )

/ F T ( T Xi ) i∈I

μ

◦F i∈I Xi

T

([T (inji )]i∈I )

/



/ FT(

i∈I

Xi )

Dependent Types and Fibred Computational Effects

47

Corollary 1. Given a monad (T, η, μ) on Set, the EM-adjunction F T U T and the families fibration p : Fam(Set) −→ Set determine a fibred adjunction model. A particularly well-behaved collection of fibred adjunction models arises from the algebraic treatment of computational effects [31], namely, from monads arising from countable Lawvere theories. These are exactly the monads on Set that are of countable rank [33, Theorem 2.8]. As discussed in [17], such monads combine easily, in terms of combining the operations and equations of the corresponding countable Lawvere theories. We recall that every countable Lawvere theory L induces a category Mod(L, Set) of models of L in Set, with an associated forgetful functor UL : Mod(L, Set) −→ Set [33]. As Set is locally countably presentable [1], UL has a left adjoint FL , inducing an equivalence of categories between Mod(L, Set) and SetTL , for TL = UL ◦ FL . Importantly for us, Mod(L, Set) is both complete and cocomplete, e.g., because TL has countable rank and Set is complete and cocomplete, see [4, Theorem 4.3.6]. Therefore, Mod(L, Set) has set-indexed products and set-indexed coproducts and we get: Corollary 2. For any countable Lawvere theory L, the adjunction FL UL and the families fibration p : Fam(Set) −→ Set determine a fibred adjunction model. Finally, we present two computationally motivated examples arising from Theorem 3 and decompositions of monads (T, η, μ) on Set into adjunctions other (−) than F T U T . In particular, we consider the continuations monad RR and S the global state monad ((−) × S) . These monads can be decomposed into the adjunctions R(−) R(−) : Setop −→ Set and (−) × S (−)S : Set −→ Set, where Setop inherits set-indexed products and set-indexed coproducts trivially from Set. Corollary 3. The adjunctions R(−) R(−) and (−) × S (−)S together with the families fibration p : Fam(Set) −→ Set determine fibred adjunction models.

4

Extending the Language with Algebraic Effects

Until now we have have not said how computational effects, such as I/O, state, exceptions, etc., arise in our language and how programmers can program with them. In this section, we make the source of computational effects explicit, by drawing ideas from the algebraic treatment of computational effects [31]. 4.1

Algebraic Effects in the Syntax

We begin by assuming we are given a collection of typed operation symbols op : (xin : I) −→ O where we call ·  I the input type and xin : I  O the output type of op. We restrict I and O to be pure value types, i.e., value types that do not contain U .

48

D. Ahman et al.

We add such operation symbols to our language by extending the syntax of computation terms: for each operation symbol op : (xin : I) −→ O, we add C algebraic operations opV (x.M ), for all C, and a generic effect genopV , typed as Γ  V :I

Γ  C

Γ, x:O[V /xin ]  M : C

C opV (x.M ) : C

Γ 

Γ  V :I Γ  genopV : F (O[V /xin ])

These operations and generic effects are in 1-to-1 correspondence, as in [30]. The operation symbols we consider can also come equipped with a collection of equations, describing their intended computational behaviour. We extend our language with the corresponding equations between computation terms. Note that we allow the output types of operation symbols to depend on input values. This additional type-dependency is useful for giving more concise representations of collections of standard, simply-typed operations. For example, we can use a boolean-indexed type family and two operation symbols lookup : (xin : Bool) −→ (if xin then Nat else Bool) update : (xin : Σx : Bool.(if x then Nat else Bool)) −→ 1 as a concise syntax for the four operations of global state with two locations: lookuptt : 1 −→ Nat

updatett : Nat −→ 1

lookupff : 1 −→ Bool

updateff : Bool −→ 1

C

In order to make the computation terms opV (x.M ) behave like algebraic operations, we extend our language with a general algebraicity equation: op : (xin : I) −→ O

Γ  V :I Γ 

Γ, x:O[V /xin ]  M : C

C K[opV (x.M )/z]

=

D opV (x.K[M/z])

Γ | z :C  K : D

:D

(∗∗)

Using (∗∗), we can easily prove equations familiar from languages with algebraic effects, e.g. the algebraicity equation for sequential composition from [34, Sect. 3.3]: C

Γ  opFA V (x.M ) to y : A inC N = opV (x.(M to y : A inC N )) : C We omit handlers of algebraic effects [32] from this paper as a full account, in which the equational properties of handlers are derived from those of homomorphism terms, i.e., from (∗∗), involves extending our language with a further computation type former A, {Vop }op  for user-defined algebras. To make handlers first-class, one extends the language with the type C  D of homomorphisms, familiar from EEC. We will report on this extension separately elsewhere. 4.2

Algebraic Effects in the Semantics

Following the definition of algebraic operations for a monad in terms of its EMalgebras [30, Sect. 6], we say that a fibred adjunction model supports algebraic operations if for all op : (xin : I) −→ O, there exist vertical morphisms ∗ opC : Σ!∗q(C) (I) (Π{!q(C) (I)}∗ (O) (π{!

∗ q(C) (I)} (O)

(π!∗∗

q(C)

(I) (U (C)))))

−→ U (C)

Dependent Types and Fibred Computational Effects

49

that are natural in C. When combining this naturality with the fact that U is a split fibred functor, it can be shown that the opC ’s are preserved by reindexing. If the given collection of operation symbols also comes with associated equations, these vertical morphisms are additionally required to satisfy these equations. The classifying model from Sect. 3.2 can be straightforwardly extended: for any Γ = x1 : A1 , . . . , xn : An and Γ  C, the vertical morphisms opΓ  C are   C opΓ  C = x1 , . . . , xn , thunk (opfst x (x.force C ((snd x ) x))) where x has type Σxin : I.Πx : O.U C. The naturality of these is proved by using the equations describing the interaction between thunking and forcing, in combination with the general algebraicity equation (∗∗) from Sect. 4.1. Finally, we note how algebraic operations can be characterised in the fibred adjunction models based on the families fibration and EM-algebras of a monad. Proposition 2. The fibred adjunction models from Sect. 3.3, based on the EM-algebras of a monad (T, η, μ) on Set, support algebraic operations if for all operations op : (xin : I) −→ O there exists a family of natural transformations  opi : (U T (−))O( ,i ) −→ U T (−) i∈I()

5

Extending the Language with General Recursion

We now show how to extend our language with general recursion, considering it as a computational effect to keep the MLTT fragment of our language effect-free. 5.1

Recursion in the Syntax

We start by extending our language with a new computation term, the fixed point operation μx : U C.M . The corresponding typing rule is given by Γ, x : U C  M : C Γ  μx : U C.M : C We also extend the language’s equational theory with unfolding of fixed points: Γ  μx : U C.M = M [thunk (μx : U C.M )/x] : C Finally, we also alter the definition of identity types IdAdisc (V, W ), restricting them to be over discrete value types, to be able to interpret this extended language in models based on continuous families of cpos. These discrete types are Adisc , Bdisc ::= Nat | Σx : Adisc .Bdisc | 1 | Πx : A.Bdisc | IdAdisc (V, W ) 5.2

Domain-Theoretic Semantics for Recursion

We build the denotational semantics for the language with recursion around the SCCompC of continuous families p : CFam(CPO) −→ CPO [18, Sect. 10.6].

50

D. Ahman et al.

Compared to [18], we use ω-complete partial orders instead of directed-complete partial orders, because the former constitute a locally countably presentable category, whereas the latter do not [1, Example 1.14(4)]; and we need local presentability for fibred adjunction models based on the algebraic treatment of computational effects. An overview of the relevant domain theory can be found in [11, 29]. We recall that the objects of CFam(CPO) are pairs (X, A) of a cpo X and a continuous functor A : X −→ CPOEP (a continuous family), treating X as a category and valued in the category of embedding-projection pairs. The morphisms (X, A) −→ (Y, B) are pairs (f, {gx }x∈|X| ) of a continuous function f : X −→ Y and a family of continuous functions {gx : A(x) −→ B(f (x))}x∈|X| , satisfying x1 X x2 =⇒ B(f (x1 X x2 ))e ◦ gx1 B(f (x2 ))A(x1 ) gx2 ◦ A(x1 X x2 )e      xn incr. ω-chain =⇒ gn xn = n B(f (xn X n xn ))e ◦ gxn ◦ A(xn X n xn )p

The dependent products and strong dependent  sums are defined by using the  cpo-indexed products X A and coproducts X A in CPO, which are given by 



|A(x)|, x1 , a1  x2 , a2 iff x1 X x2 and A(x1 X x2 )e (a1 )  a2    X A = {f : X −→ X A | fst ◦ f = idX }, f1  f2 iff ∀x ∈ |X|. f1 (x)  X A f2 (x)

X

A=





x∈|X|

For identity types, we require A to be a continuous family of discrete cpos, matching the changes we made in the syntax of our language, and then define  ∗   Id(X,A) = {π(X,A) (X, A)}, x, a, a  → { | a=a } 1  The discreteness of A is necessary for x, a, a  → { | a=a } 1 to constitute a continuous functor: if A would not be discrete, then from x1 , a1 , a1   x2 , a2 , a2  and a1 = a1 it would not follow that a2 = a2 , and we need for  vice versa, which defining the embedding-projection pair between { | a1 =a } 1 and { | a2 =a } 1. 1 2 For modeling computation terms involving recursion, we assume a CPOenriched monad (T, η, μ) on CPO, such that its EM-algebras are pointed and the morphisms between them are strict; or equivalently, we assume that the given monad supports a least zero-ary algebraic operation, in the sense of [30, Sect. 6]. For modeling computational Σ-types, we further assume that CPOT has reflexive coequalizers. We can then model our computation types in the split fibration q : CFam(CPOT ) −→ CPO, defined analogously to p above. Monads satisfying these conditions arise naturally from the algebraic treatment of computational effects: from CPO-enriched countable Lawvere theories [17] with a least constant. The dependent products and sums in q, with respect to p, are defined using the cpo-indexed products and coproducts in CPOT . First, the cpo-indexed prod uct X C in CPOT is directly inherited from CPO, being defined on the carT ◦ C). On the other hand, analogously toSect. 3.3, the cpo-indexed rier X (U T coproduct X C cannot be defined simply by taking X (U ◦ C) as the carrier. Instead, we construct X C as the reflexive coequalizer for a diagram similar to the one used in Sect. 3.3, with the difference that here we would use the free algebras over cpo-indexed coproducts rather than over set-indexed coproducts.

Dependent Types and Fibred Computational Effects

51

Despite having a more complex categorical definition, the cpo-indexed as those in CPO. We coproducts in CPOT have the same universal property  from applying first recall Jacobs’s remark [18, Sect. 10.6] that X A results  the Grothendieck construction to A. By a result of Gray [12], X A can also be understood as the oplax colimit of A in CPO. In CPOT , the situation is analogous:  Proposition 3. X C is the oplax colimit of C : X −→ (CPOT )EP in CPOT . Similarly to Sect. 3.3, the split fibred adjunction F U : q −→ p is defined from F T U T by post-composition, i.e., by setting F (X, A) = (X, F T ◦A), where F T ◦ A is a continuous functor because of the CPO-enrichment of F T and the limit-colimit coincidences in CPOEP and (CPOT )EP . U is defined analogously. We interpret our language as discussed in Sect. 3.2, except for recursion:   Γ ; μx : U C.M  = idΓ  , {x ∈ |1| → μ(gγ )}γ∈|Γ | : (Γ , γ → 1) −→ (Γ , U T ◦ Γ ; C)

where we use the least fixed points of the family of continuous functions {gγ : U T (Γ ; C)(γ) −→ U T (Γ ; C)(γ)}γ∈|Γ | , determined by a vertical morphism (Γ , U T ◦ Γ ; C) −→ (Γ , U T ◦ Γ ; C) we derive from Γ, x : U C; M . The least fixed points of gγ ’s are guaranteed to exist because our assumptions about CPOT make every U T ◦ Γ ; C into a continuous family of pointed cpos. Theorem 4. Given a monad (T, η, μ) on CPO satisfying the conditions given in this section, the fibred adjunction model built from p : CFam(CPO) −→ CPO and F T U T is a model of the equational theory extended with fixed point unfolding. Finally, we note that the other obvious candidate cod : CPO→ −→ CPO, even if made split [8,14], is not a SCCompC, because of [18, Theorem 10.5.5] and: Proposition 4. CPO is not locally Cartesian closed. In particular, the condition that every base change functor has to have a right adjoint fails because some of these functors do not preserve all colimits, e.g., given a non-empty cpo X, the pullback of the epimorphism n → n : N= −→ Nω in CPO/Nω along the constant map x → ω : X −→ Nω is not an epimorphism.

6

Conclusions and Future Work

We addressed the problem of finding a mathematically natural combination of dependent types and computational effects. We were motivated by: (i) the success similar foundations have had in driving the study of computational effects in the simply-typed setting; and (ii) the success of dependently-typed programming in generating a number of concrete attempts to combine dependent types with computational effects. Our solution is mathematically natural, combining comprehension categories, arising from the semantics of dependent types, with adjunctions, arising from the semantics of computational effects. It is also

52

D. Ahman et al.

general, covering a variety of algebraic and non-algebraic effects, and can be extended to accommodate general recursion. For future work, a natural next step is to investigate operational semantics, leading towards an implementation. We are also working on a fibred generalisation of Atkey’s parametrised notions of computation [2], aiming at a semantic account of the effects in Idris [5,6].

References 1. Adamek, J., Rosicky, J.: Locally Presentable and Accessible Categories. London Mathematical Society Lecture Note Series. Cambridge University Press, Cambridge (1994) 2. Atkey, R.: Algebras for parameterised monads. In: Kurz, A., Lenisa, M., Tarlecki, A. (eds.) CALCO 2009. LNCS, vol. 5728, pp. 3–17. Springer, Heidelberg (2009) 3. Benton, P.N.: A mixed linear and non-linear logic: proofs, terms and models. In: Pacholski, L., Tiuryn, J. (eds.) CSL 1994. LNCS, vol. 933. Springer, Heidelberg (1995) 4. Borceux, F.: Handbook of Categorical Algebra. Categories and Structures, vol. 2. Cambridge University Press, Cambridge (1994) 5. Brady, E.: Programming and reasoning with algebraic effects and dependent types. In: Morrisett, G., Uustalu, T. (eds.) Proceedings of 18th International Conference on Functional Programming, ICFP 2013, pp. 133–144. ACM (2013) 6. Brady, E.: Resource-dependent algebraic effects. In: Hage, J., McCarthy, J. (eds.) TFP 2014. LNCS, vol. 8843, pp. 18–33. Springer, Heidelberg (2015) 7. Casinghino, C., Sj¨ obergberg, V., Weirich, S.: Combining proofs and programs in a dependently typed language. In: Sewell, P. (ed.) Proceedings of 41st Annual Symposium on Principles of Programming Languages, POPL 2014, pp. 33–45. ACM (2014) 8. Clairambault, P., Dybjer, P.: The biequivalence of locally cartesian closed categories and martin-l¨ of type theories. In: Ong, L. (ed.) Typed Lambda Calculi and Applications. LNCS, vol. 6690, pp. 91–106. Springer, Heidelberg (2011) 9. Egger, J., Møgelberg, R.E., Simpson, A.: The enriched effect calculus: syntax and semantics. J. Log. Comput. 24(3), 615–654 (2014) 10. Ghani, N., Johann, P., Fumex, C.: Indexed induction and coinduction, fibrationally. Logical Methods Comput. Sci. 9(3:6), 1–31 (2013). doi:10.2168/LMCS-9(3:6)2013 11. Gierz, G., Hofmann, K.H., Keimel, K., Lawson, J.D., Mislove, M., Scott, D.S.: Continuous Lattices and Domains. Cambridge University Press, Cambridge (2003) 12. Gray, J.W.: The categorical comprehension scheme. In: Gray, J.W. (ed.) Category Theory, Homology Theory and Their Applications III. Lecture Notes in Mathematics, vol. 99, pp. 242–312. Springer, Heidelberg (1969) 13. Hancock, P., Setzer, A.: Interactive programs in dependent type theory. In: Clote, P.G., Schwichtenberg, H. (eds.) CSL 2000. LNCS, vol. 1862, pp. 317–331. Springer, Heidelberg (2000) 14. Hofmann, M.: On the interpretation of type theory in locally cartesian closed. In: Pacholski, L., Tiuryn, J. (eds.) CSL 1994. LNCS, vol. 933, pp. 427–441. Springer, Heidelberg (1995) 15. Hofmann, M.: Syntax and semantics of dependent types. In: Pitts, A.M., Dybjer, P. (eds.) Semantics and Logics of Computation, pp. 79–130. Cambridge University Press, Cambridge (1997)

Dependent Types and Fibred Computational Effects

53

16. Hutton, G., Meijer, E.: Monadic parsing in Haskell. J. Funct. Program. 8(4), 437– 444 (1998) 17. Hyland, M., Plotkin, G., Power, J.: Combining effects: sum and tensor. Theor. Comput. Sci. 357(1–3), 70–99 (2006) 18. Jacobs, B.: Categorical Logic and Type Theory. Studies in Logic and the Foundations of Mathematics, vol. 141. North Holland, Amsterdam (1999) 19. Krishnaswami, N.R., Pradic, P., Benton, N.: Integrating linear and dependent types. In: Walker, D. (ed.) Proceedings of 42nd Annual Symposium on Principles of Programming Languages, POPL 2015, pp. 17–30. ACM (2015) 20. Levy, P.B.: Call-By-Push-Value: A Functional/Imperative Synthesis. Semantics Structures in Computation, vol. 2. Springer, Heidelberg (2004) 21. Linton, F.: Coequalizers in categories of algebras. In: Eckmann, B. (ed.) Seminar on Triples and Categorical Homology Theory. Lecture Notes in Mathematics, vol. 80, pp. 75–90. Springer, Heidelberg (1969) 22. Martin-L¨ of, P.: An intuitionisitc theory of types, predicative part. In: Rose, E., Shepherdson, J.C. (eds.) Proceedings of Logic Colloquium 1973, pp. 73–118. North-Holland (1975) 23. McBride, C.: Functional pearl: Kleisli arrows of outrageous fortune. J. Funct. Program. (To appear) 24. Moggi, E.: Computational lambda-calculus and monads. In: Parikh, R. (ed.) Proceedings of 4th Annual Symposium on Logic in Computer Science, LICS 1989, pp. 14–23. IEEE (1989) 25. Moggi, E.: Notions of computation and monads. Inf. Comput. 93(1), 55–92 (1991) 26. Nanevski, A., Morrisett, G., Birkedal, L.: Hoare type theory, polymorphism and separation. J. Funct. Program. 18(5–6), 865–911 (2008) 27. Pitts, A.M.: Categorical logic. In: Abramsky, S., Gabbay, D.M., Maibaum, T.S.E. (eds.) Handbook of Logic in Computer Science. Algebraic and Logical Structures, vol. 5, pp. 39–128. Oxford University Press, Oxford (2000). chap. 2 28. Pitts, A.M., Matthiesen, J., Derikx, J.: A dependent type theory with abstractable names. In: Mackie, I., Ayala-Rincon, M. (eds.) Proceedings of 9th Workshop on Logical and Semantic Frameworks, with Applications, LSFA 2014. ENTCS, vol. 312, pp. 19–50. Elsevier (2015) 29. Plotkin, G.: Pisa notes (on domain theory) (1983) 30. Plotkin, G., Power, J.: Semantics for algebraic operations. In: Brookes, S., Mislove, M. (eds.) Proceedings of 17th Conference on the Mathematical Foundations of Programming Semantics, MFPS XVII. ENTCS, vol. 45, pp. 332–345. Elsevier (2001) 31. Plotkin, G., Power, J.: Notions of computation determine monads. In: Nielsen, M., Engberg, U. (eds.) FOSSACS 2002. LNCS, vol. 2303, pp. 342–356. Springer, Heidelberg (2002) 32. Plotkin, G.D., Pretnar, M.: Handling algebraic effects. Logical Methods Comput. Sci. 9(4:23), 1–36 (2013). doi:10.2168/LMCS-9(4:23)2013 33. Power, J.: Countable Lawvere theories and computational effects. In: Seda, A.K., Hurley, T., Schellekens, M., an Airchinnigh, M.M., Strong, G.(eds.) Proceedings of 3rd Irish Conference on the Mathematical Foundations of Computer Science and Information Technology, MFCSIT 2004. ENTCS, vol. 161, pp. 59–71. Elsevier (2006) 34. Pretnar, M.: The Logic and Handling of Algebraic Effects. Ph.D. thesis, School of Informatics, University of Edinburgh (2010)

54

D. Ahman et al.

35. Staton, S.: Instances of computational effects: an algebraic perspective. In: Kupferman, O. (ed.) Proceedings of 28th Annual Symposium on Logic in Computer Science, LICS 2013, pp. 519–519. IEEE, June 2013 36. Streicher, T.: Semantics of Type Theory. Birkh¨ auser, Boston (1991) 37. Swamy, N., Weinberger, J., Schlesinger, C., Chen, J., Livshits, B.: Verifying higherorder programs with the Dijkstra monad. In: Flanagan, C. (ed.) Proceedings of 34th Conference on Programming Language Design and Implementation, PLDI 2013, pp. 387–398. ACM (2013) 38. V´ ak´ ar, M.: A categorical semantics for linear logical frameworks. In: Pitts, A. (ed.) FOSSACS 2015. LNCS, vol. 9034, pp. 102–116. Springer, Heidelberg (2015) 39. Warren, M.A.: Homotopy Theoretic Aspects of Constructive Type Theory. Ph.D. thesis, Department of Philosophy, Carnegie Mellon University (2008)

Game Semantics for Bounded Polymorphism James Laird(B) Department of Computer Science, University of Bath, Bath, UK [email protected]

Abstract. We describe a denotational, intensional semantics for programs with polymorphic types with bounded quantification, in which phenomena such as inheritance between stateful objects may be represented and studied. Our model is developed from a game semantics for unbounded polymorphism, by establishing dinaturality properties of generic strategies, and using them to give a new construction for interpreting subtyping constraints and bounded quantification. We use this construction to give a denotational semantics for a programming language with general references and an expressive polymorphic typing system. We show that full abstraction fails in general in this model, but that it holds for all terms at a rich collection of bounded types.

1

Introduction

By combining subtype and parametric polymorphism, type systems with bounded quantification increase the expressive power of both: they may be used to write programs which are generic, but range over a constrained set of types (a program of type ∀X ≤ S.T may be instantiated only with a subtype of S). They have been used to develop formal theories of key aspects of object oriented languages such as inheritance [6]. Our aim is to develop an intensional denotational semantics for subtyping and bounded polymorphism — i.e. a formal semantic account of the constraints on behaviour which can be expressed in such a system. This allows for models which combine bounded polymorphism with computational effects (in particular, state) and, potentially, for semantics-based subtyping theories which capture aspects of program behaviour. Previous denotational models of bounded quantification have been based on an extensional interpretation of polymorphism — for instance by interpreting subtyping as inclusion of partial equivalence relations [5] or as a relation between games [8], so that bounded quantification corresponds to a product or intersection over all instances satisfying the bound. This has generated valuable insights into the theory of subtyping and polymorphism, but it is not clear how it may be extended to include computational effects (for example). We take an alternative, more intrinsically intensional approach by constructing an interpretation of bounded quantification without first specifying a subtyping relation, beyond the requirement that it gives rise to coercion morphisms between objects: we will subsequently show how to interpret subtyping judgements as such coercions. c Springer-Verlag Berlin Heidelberg 2016  B. Jacobs and C. L¨ oding (Eds.): FOSSACS 2016, LNCS 9634, pp. 55–70, 2016. DOI: 10.1007/978-3-662-49630-5 4

56

J. Laird

Technically, our model is based on the game semantic framework for generic call-by-name polymorphism described in [15] and reviewed and updated in Sect. 2, which develops earlier notions of variable game [1, 12] with a new interpretation of quantification as a relation between question and answer moves. In Sect. 3, we extend this framework by showing that context games correspond to mixed variance functors on games, and generic strategies to (composable) dinatural transformations between them. In Sect. 4, we use these properties to construct an interpretation of subtyping environments and bounded quantification. The basic idea is to represent a bounded quantification ∀X ≤ S.T (X − , X + ) as an unbounded quantification ∀X.T (S × X, X), so that negative occurrences of type variables may be coerced into their bounding types by right projection: defining the composition of strategies with free, constrained variables requires precisely the dinaturality properties established in the previous section. In Sect. 5 we give a semantics for a stateful programming language with a second-order typing system based on Cardelli and Wegner’s Bounded Fun [6], and its extensively studied fragment, System F≤ (“F-sub”) [7, 9] (an extension of the second-order λ-calculus, System F [11,17] with bounded quantification). However, these are pure type theories: the language we study is (in essence) a conservative extension of the metalanguage with general references, and its games model, originally defined (and proposed as a semantic basis for object oriented programs) by Abramsky, Honda and McCusker [3]. Finally, we show that full abstraction does not hold in our model (arguably, due to an expressive deficit in the language and typing system) but holds for a significant fragment determined by a simple constraint on types.

2

Second Order Game Semantics

Our games model is based on a semantics of unbounded “System F - style” polymorphism which is essentially equivalent to that defined in [13, 15], with some minor differences in the representation of games (up to these, our model of bounded quantification is a conservative extension). Here we recall and modify the key definitions, referring to [15] for details and proofs. A “n-context game” as defined in [13] is a Hyland-Ong arena — a set of moves partitioned between two players, with a binary enabling relation defining a directed, bipartite acyclic graph — with a further partitioning of moves into sets of questions, answers, and i-holes for i ∈ {1, . . . , n} into which other arenas may be instantiated, and a scoped question answer/relation — a ternary relation on moves — q m a means a can answer q within the scope of m — such that Player questions are scoped by Player moves and answered by Opponent moves, and Opponent moves are scoped by Opponent moves and answered by Player moves. In the following, we will work within a universal set of moves in which enabling, labelling of moves and the question/answer relation are determined by a fixed representation — so a context arena may be specified simply by giving a set of moves within this universe (although this is not essential for our model of bounded quantification, it facilitates some constructions).

Game Semantics for Bounded Polymorphism

57

Definition 1. Let A be the set N ∪ {l, r} ∪ {∀i | i ∈ N}. Define the universal set of moves U to be {w •n | w ∈ A∗ ∧ n ∈ N\{0}} — i.e. a move is a word over A followed by a terminal symbol •n for some n ≥ 1. Say that a sequence w ∈ A∗ is positive if it contains an even number of occurrences of the symbol l, and negative otherwise. If w is positive, then w•i is an Opponent move, otherwise it is a Player move. (Given X ⊆ U, X + is the set of Opponent moves and X − the set of Player moves in X.) Definition 2. We define the enabling relation ⊆ U × U: m n if there exist sequences u, v, w with v, w not containing l, such that m = u · v and n = u · lw. This is evidently a directed, bipartite acyclic graph on U. Source nodes (Opponent moves containing no occurrence of l) are called initial moves. Definition 3. Moves are partitioned into questions, answers, and i-holes as follows: – If m = v•i where v contains no occurrences of ∀i , then m is an i-hole move. – Otherwise, m = u∀i · v•i , for some unique v which contains no occurrences of ∀i . Then m is a question if v is negative, and an answer if v is positive. The (ternary) scoped question/answer relation on moves is defined as follows: – q m a if there exists t, u, v, w such that u contains no occurrences of l and u, v, w contain no occurrences of ∀i , and m = t∀i · u, q = t∀i · v•i and a = ∀i · w•i . Definition 4. A game is a subset A ⊆ U such that if u•i ∈ A is a hole move then u contains no occurrences of ∀j for j ≤ i, and if u m ∈ A then m = u•i . A is a n-context game if HA (i) (the set of i-hole moves in A) is empty for i > n. 2.1

Examples

Given a n-context game A and element a ∈ A, we write a.A for {am | m ∈ A}. – We denote the empty game (the empty set of moves) by 1. – For each i ≤ n, {•i } (or just •i ) is the n-context game containing the single type variable. (Opponent hole) move •i , corresponding to a free – For any indexing set I ⊆ N, the disjoint union i∈I i.A corresponds to an indexed product or record type. – Let A ⇒ B denote the game l.A ∪ r.B (the disjoint union of A and B, with the initial moves of B becoming enablers for the initial moves of A, and the moves of A switched between Player and Opponent. – If A is a n-context game, ∀n .A is a n − 1-context game; the Opponent i-hole moves of A become questions, scoped by the O-initial moves of A, and the Player i-holes of A become their answers.

58

J. Laird

In this way, we can interpret an (unbounded) second-order type (in which variables and their binders are represented as occurrences of • and ∀ with de Bruijnstyle indices) directly as a game in which the moves correspond to the paths through the syntax tree of the type from the root to a leaf node (bound or free variable). For example, suppose A and B are n-context games. The game ∀n+1 ((A ⇒ •n+1 ) ⇒ (B ⇒ •n+1 ) ⇒ •n+1 ) (corresponding to the System F type ∀Xn+1 .(A → Xn+1 ) ⇒ (B → Xn+1 ) → Xn+1 ) consists of the initial Opponent question ∀n+1 rr•n+1 , which enables and scopes its two (Player) answers ∀n+1 rlr•n+1 and ∀n+1 lr•n+1 , and the moves ∀n+1 ll.A and ∀n+1 rll.B, of which the moves of the form ∀n+1 llm and ∀n+1 rllm where m is initial in A or B respectively, are enabled by ∀n+1 rlr•n+1 and ∀n+1 lr•n+1 , respectively. In other words, this is the lifted sum game [16], used to define a computational monad, and thus a call-by-value semantics in [2], which we will extend with bounded polymorphism. 2.2

Legal Sequences and Strategies

Interaction in the game may be seen as alternately choosing leaf nodes of this syntax tree, according to rules we now describe. Define the pending question prefix of a sequence of moves t over a 0-context arena A (if any): – pending(sq) = sq, if q is a question, – pending(sa) = pending(s ), if a is an answer, and pending(s) = s q. Definition 5. A legal sequence t on A is a finite sequence of moves in A, alternating between Opponent and Player moves, equipped with a unique justification pointer from each non-initial move to a preceding move which enables it, such that: – If sqs a t, where q is the pending question in sqs then there exists a move m in s which hereditarily justifies both q and a, such that q m a (the bracketing condition). t is single threaded if it contains at most one initial move, in which case any pointers from moves enabled by (and therefore justified by) the initial move are considered implicit. A strategy for a 0-context game A is a non-empty, even-prefix-closed, evenbranching set of single-threaded legal sequences on A. For each n, we define a category G(n) in which objects are n-context games and morphisms from A to B are strategies on ∀1 . . . ∀n (A ⇒ B) (we will also write G(0) as just G). To define composition in G(n), we use the following generalised notion of restriction on justified sequences: given a partial function f : A → B we may derive a function on sequences over B — f ∗ : U ∗ → U ∗ applies f pointwise to the moves on which it is defined, and omits moves on which it is not defined. This extends to justified sequences: f (n) points to f (m) in f ∗ (s) iff f (m) and f (n) are both defined and n points to m in s. Where f is evident from the context, we shall write sB for f ∗ (s).

Game Semantics for Bounded Polymorphism

59

Let σ † be the least set of legal sequences containing σ and closed under interleaving. The composition of σ : A → B with τ : B → C is defined: σ; τ = {s ∈ L∀(A⇒C) | ∃t ∈ A + B + C.t∀(A ⇒ B) ∈ σ † ∧ t∀(B ⇒ C) ∈ τ ∧ s = t  ∀(A ⇒ C)}. The identity on A is the set of copycat sequences {t ∈ L∀(A⇒A) | ∀s E t.sA+ = sA− }. For any n-context games A and B, A × B = 0.A ∪ 1.B is a Cartesian product of A and B in G(n), and A ⇒ B is an internal hom (so G(n) is a Cartesian closed category). For each n, if A is a n-context game, and B is a n + 1 context game, there is an evident isomorphism between the arenas ∀n+1 (A ⇒ B) and A ⇒ ∀n+1 B, yielding an evident natural correspondence between G(n + 1)(Jn+1 (A), B) and G(n)(A, ∀n+1 .B) — i.e. the inclusion Jn+1 : G(n) → G(n + 1) has ∀n+1 as its right adjoint [15]. 2.3

Instantiation

Let A be a n-context game, and B and C be i-context games, with i ≤ n. The instantiation of B and C into the (negative and positive) i-holes of A (respectively) is defined A(B, C)i =: A\HA (i) ∪ {w · m | w•i ∈ HA (i)− ∧ m ∈ B} ∪ {w · m | w•i ∈ HA (i)+ ∧ m ∈ C}. This operation behaves as one would expect for a substitution1 : e.g. if A = •1 ⇒ •1 ⇒ •1 then A(B, C)1 = B ⇒ B ⇒ C. Moreover, each morphism in G(n)(A, B) corresponds to a generic family of morphisms from A(C, C) to B(C, C) for each n-context game C. These are defined (as in [15]) using the notion of instantiation of a game into strategy by playing copycat between the arenas plugged into the holes. Suppose t is a legal sequence on ∀(A(C, C)i ⇒ B(C, C)i ), such that t∀(A ⇒ B) is a legal sequence, and sa t∀(A ⇒ B), where a = v•i is a Player answer in ∀(A ⇒ B) corresponding to a i-hole move in A ⇒ B. As t is well bracketed, pending(s) = s q where q = u•i is an Opponent question in ∀(A ⇒ B) corresponding to an i-hole move in A ⇒ B. So we may define t(pending(s), sa) to be the projection of t onto ∀(C ⇒ C) of moves which are suffixes of u and v (and hereditarily justified by q and a, respectively). Definition 6. Given a strategy σ : A → B, let σ[C]i : A(C, C)i → B(C, C)i be the set of sequences t ∈ L∀(A(C,C)i ⇒B(C,C)i ) such that t∀(A ⇒ B) ∈ σ and if sa E t, where a is a Player i-hole move in A ⇒ B, then t(pending(s), sa)) ∈ idC . As shown in [15], the G(n) may be organised as an indexed cartesian closed category of context games, in which the Jn and ∀i form an indexed adjunction. Here we note the following result from loc. cit.. Lemma 1. [C]i is an endofunctor on G(n), such that [C]i · [D]i = [B(C, C)i ]i . 1

The condition that i-hole moves contain no occurrence of ∀j for j ≤ i prevents “capture” of the holes.

60

3

J. Laird

Generic Strategies as Dinatural Transformations

In addition to the structure defined in [15], instantiation also acts on strategies as a (mixed variance) functor (which we will require in order to define our semantics of bounded quantification) in the following sense: + (i), let pw•i be the partial Definition 7. Given an Opponent hole w•i ∈ HA projection on moves sending ∀lw · n to ∀ln and ∀rw · n to ∀rn, so that it projects − (i) is a a move in ∀(A(B, C)i ⇒ A(B  , C  )i ) into ∀(C ⇒ C  ). If m ∈ HA Player hole move, then let pm be the corresponding projection from ∀(A(B, C)i ⇒ A(B  , C  )i ) into ∀(B  ⇒ B). Now, given morphisms σ : B  → B, τ : C → C  in G(n), we may define A(σ, τ )i : A(B, C) → A(B  , C  ) to be the set of even-length sequences t ∈ L∀(A(B,C)⇒A(B  ,C  )) such that for all even prefixes s t, s∀(A ⇒ A) ∈ idA and for any hole move m ∈ Hi+ (A), p∗m (s) ∈ σ † and for any m ∈ Hn− (A), p∗m (s) ∈ τ † .

Proposition 1. For any n-context game A, A( , )i is a functor from G(n)op × G(n) to G(n). One might hope that for any generic n-context strategy σ : A → B, the family {σ[C] : A(C, C)i → B(C, C)i } is a dinatural transformation [4] from A( , )i to B( , )i — i.e. that for any morphism τ : C → D, the following diagram commutes: A(C, C)

σ[C]

A(τ,C) q8

q qqq A(D, C) MMA(C,τ MMM ) & A(D, D)

σ[D]

/ B(C, C) NND(C,τ NNN ) & B(C, D) A(τ,D) p8 p ppp / B(D, D)

However, this is not the case: Proposition 2. σ[ ]i is not a dinatural transformation from A( , )i to B( , )i in general. Proof. Take, for example, A = ∀1 (•1 ⇒ •1 ), B = •1 ⇒ •1 and σ : ∀1 (•1 ⇒ •1 ) → (•1 → •1 ) to be the counit of the adjunction J1  ∀1 (i.e. the copycat strategy between A and B). At any game, ⊥ is the strategy containing only the empty sequence. Then A(⊥, 1); σ[1]; B(1, ⊥) = ⊥, but A(•1 , ⊥); σ[•1 ]; B(⊥, •i ) = ⊥ (it contains a response to the initial move). The generic strategies (in Hughes’ model of System F) which do correspond to dinatural transaformations are characterized in [10], to which we refer for further discussion. Here we take an alternative approach by showing that σ[ ]i is dinatural with respect to a subcategory of linear morphisms.

Game Semantics for Bounded Polymorphism

61

Definition 8. A morphism σ : A → B is linear if for every initial move m in B there is an initial move m in A such that msm ∈ σ if and only if s = ε (so every non-empty sequence s ∈ σ contains exactly one initial move from A). It is easy to see that the composition of linear morphisms is linear, and so for each n we may form a subcategory G L (n) of G(n) consisting of n-context games and linear morphisms, with an inclusion JL : G L (n) → G(n). Note that × is a cartesian product in G L (n). Proposition 3. σ[ ]i is a dinatural transformation from A · JL to B · JL . In other words, for any linear morphism τ : C → D the dinaturality hexagon does commute. We prove this by defining the instantiation of τ directly into σ, i.e. σ[τ ]i : A(B, B)i → D(C, C)i by replacing the identity on B with τ in Definition 6 and showing that A(C, τ )i ; σ[C]i ; D(τ, C)i and A(τ, B)i ; σ[B]i ; D(B, τ )i are both equal to σ[τ ]i . Although the composition of dinatural transformations is not, in general, dinatural [4], by Proposition 1, those which arise by instantiation of generic strategies do compose.

4

Semantics of Bounded Quantification

We may now use the functors and dinatural transformations derived from instantiation to represent bounded quantification. A subtyping constraint of the form Xi ≤ T , where Xi is a type variable, corresponds to the ability to subsume any term of type Xi into the type T . Thus, in an environment where such a constraint holds, negative occurences of Xi may be represented as the game [[T ]] × •i , from which there is a canonical coercion (left projection) into [[T ]]. On the other hand, positive occurrences of Xi are simply represented as •i (we can’t, for example, decompose Xi into the bound T and an “unknown” extension). The difficulty is to define a notion of composition between morphisms by “unifying” this different treatment of positive and negative occurrences2 . The key to doing so is the following operation: Definition 9. Let A be an n − 1-context game. Given σ : B(•n , A × •n )n →  : B(A × •n , A × •n )n → C(A × •n , A × •n )n as follows: C(A × •n , •n )n , define σ (δ : A → A × A = A, A). B(A × •n , A × •n )n

B(Aוn ,δוn )n

−→

C(A × A × •n , A × •n )n

B(A × •n , A × A × •n )n

C(δוn ,Aוn )n

−→

σ[Aוn ]n

−→

C(A × •n , A × •n )n

We illustrate by describing a setting corresponding to an environment containing a single constrained type variable. Given a n − 1-context game A, we may define a category GA in which objects are n-context games and morphisms from B to C 2

Representing negative and positive occurrences of Xi as A × •i leads to a simple interpretation of composition, but to the failure of bounded universal quantification to be antitone in type bounds — i.e. we get a model of “kernel F≤ ” [6].

62

J. Laird

are the morphisms from B(•n , A × •n ) to C(A × •n , •n ) in G. The composition of  : B(•n , A× σ ∈ GA (B, C) with τ ∈ GA (C, D) is the composition of B(πr , Aו); σ •n ) → C(Aוn , Aוn ) with τ; D(Aוn , πr ) : C(Aוn , Aוn ) → D(Aוn , •n ) in G. The identity on B in GA is B(πr , πr ). Lemma 2. The operation () satisfies the following properties:  1. B(π r , πr ) = B(A × •, A × •). ; C(A × •, πr ). 2. For σ : B(•, A × •) → C(A × •, •), σ = B(πr , A × •); σ σ ; τ; D(A × •, πr ). 3. For τ : C(•, A × •) → D(A × •, •), σ ; τ = B(πr , A × •); Proposition 4. GA is a well-defined category. Proof. That B(πr , πr ) is an identity follows from Lemma 2 (1). For associativity of composition, suppose ρ ∈ GA (B, C), σ ∈ GA (C, D) and τ ∈ GA (D, E). Then ρ; σ ; C(A × •, πr ); τ; D(πr , A × •, πr ). By (ρ; σ); τ = B(πr , A × •); B(πr , A × •); ; τ; E(πr , A × •, πr ), which is Lemma 2 (3), this is equal to B(πr , A × •); ρ; σ similarly equal to ρ; (σ; τ ). We also have operations on GA corresponding to bounded quantification and instantiation: let ∀A C = ∀.C(A × •, •). Then for any 0-context game B, there is an evident isomorphism of arenas between ∀(B ⇒ C(A × •, •)) and D ⇒ ∀A C yielding: Proposition 5. The inclusion JA : G → GA is left adjoint to ∀A : GA → G. Definition 10. Given a 0-context arena D, and a linear morphism c : D → A, representing a coercion of D into A, we define the bounded instantiation functor sending σ : B → C to σ{c, D} : B(D, D) → C(D, D): B(D, D)

B(D,c,D )

−→

σ[D]

B(D, A × D) −→ C(D × A, D)

C(c,D ,D)

−→

C(D, D)

Proposition 6. {c, D} is a (Cartesian closed) functor from GA to G. We iterate these constructions to define categories of context games for representing types with multiple bounded free variables. Definition 11. A bounding sequence of context games A1 , . . . , An is one in which each Ai is a i−1-context game. Given such a sequence, we define categories G(n)A1 ,...,Ai (and its subcategory G(n)L A1 ,...,Ai ) for each i ≤ n, in which objects are n-context games and: – G(n)ε = G(n). – In G(n)A1 ,...,Ai+1 , morphisms from B to C are the morphisms from B(•i , Ai+1 × •i+1 )i+1 to C(Ai+1 × •i+1 , •i+1 ) in G(n)A1 ,...,Ai , with the composition of σ : B → C with τ : C → D defined to be the composition of i with τi ; C(Ai+1 × •i , πr ) in G(n)A1 ,...,Ai . The idenB(πr , Ai+1 × •i+1 )i+1 ; σ tity on B is B(πr , πr )1 , . . . , (πr , πr )i+1 .

Game Semantics for Bounded Polymorphism

63

Proof (by induction) that each GA1 ,...,Ai is a (Cartesian closed) category follows the proof of Proposition 4 above. For any bounding sequence A1 , . . . , An , B we B have a right adjoint ∀B n+1 : to the inclusion functor Jn+1 : G(n)A1 ,...,An → G(n + 1)A1 ,...,An ,B . Definition 12. For any i < n, i-context game C and c ∈ G(n)L A1 ,...,Ai−1 (C, Ai ), the instantiation functor {c, C}i from G(n)A1 ,...,Ai to G(n)A1 ,...,Ai−1 sends D to D(C, C)n+1 and σ : D → E to D(C, c, C)i ; σ[C]i ; E(c, C, C)i . This lifts to a functor from G(n + 1)A1 ,...,An to G(n)A1 ,...,Ai−1 ,Ai+1 (Ai ,Ai ),...,Aj (Ai ,Ai ) for each n ≥ j > i, since e.g. D(Aj × •j , •j )(C, C)i = D(C, C)i (Aj (C, C)i × •j , •j ) Instantiation preserves cartesian closed structure, and also commutes with the second-order structure, in the following sense: A(C,C)

Proposition 7. {c, C}i · ∀A j = ∀j

5

· {c, C}i .

A Stateful Language with Bounded Quantification

So far, we have described subtyping constraints on variables, and bounded quantification, without explicitly defining or discussing the semantics of the subtyping relation on games, beyond the assumption that it yields a coercing linear morphism from a subtype to its supertype. In this sense, our semantics is somewhat independent of any particular notion of subtyping, which can be supplied either intensionally, by a formal system of derivations for subtyping judgments with their interpretation as coercing morphisms, or extensionally, arising from a relation on objects in our category with a corresponding notion of coercion (for example, Chroboczek’s notion of subtyping for games [8]). Taking the former approach, our interpretation of subtyping constraints, bounded quantification and instantiation is sufficient to model typing systems such as Cardelli and Wegner’s Fun [6] or System F≤ [7]. However, since we also wish to describe the behaviour of (possibly side effecting) programs in this context, we define a stateful, call-by-value metalanguage, L≤ , with such a typing system. (The main difference with respect to F≤ as a typing system is the inclusion of explicit object (i.e. record) types: although record types can be encoded in F≤ , the semantics of this encoding in our model is not faithful to the direct interpretation of such types as products). In essence, L≤ and its model are an extension with subtyping and bounded quantification of the metalanguage with general references, L, defined with its games model by Abramsky, Honda and McCusker [3] (from which there is a simple translation into L≤ ). Raw types of L≤ are generated by the grammar: S ::= unit | nat | S → O | ∀X ≤ O.O

O ::= X | [l1 : S1 , . . . , ln : Sn ]

(where the li are drawn from an unbounded collection of labels). Type variables are restricted to range over the (concrete and variable) object types O, which are assigned to records containing (possibly side-effecting) methods typed according

64

J. Laird

to the values they may return. We can wrap any method type S in an object type [S] with a single field, for which we omit the label (corresponding to the lifting operation of the computational λ-calculus), and “thunk” an object type O as the method type O = unit → O. Judgments Σ T (T is a well-formed type with free variables in Σ) are derived via the rules: Table 1. Subtyping judgments for L≤

Table 2. Typing judgments for L≤

A subtyping context is a sequence of subtyping assumptions X1 ≤ O1 , . . . , Xn ≤ On such that X1 , . . . , Xi−1 Oi for each 1 ≤ i ≤ n. |X1 ≤ O1 , . . . , Xn ≤ On | is the sequence X1 , . . . , Xn . Subtyping judgments Θ T ≤ T  — where Θ = X1 ≤ O1 , . . . , Xn ≤ On is a subtyping context and |Θ| T and |Θ| T  — are derived according to the rules in Table 1. Note that there is a ≤-greatest object type — the record type with no labels, for which we will write . So (as in F≤ ) we can represent unbounded quantification ∀X.O as ∀X ≤ .O. There is no greatest method type, however. Nor do we capture inclusion of values in the subtyping relation, but leave it as a possible extension. The point is that the distinction between subtyping as inclusion of values, versus extension of records, which is blurred in the pure type theory F≤ , becomes necessarily more significant in a setting with side effects. We define the type var[S] of variables storing values of type S to be the object type with two methods, [get : S, set : S → [unit]].

Game Semantics for Bounded Polymorphism

65

Typing judgments Θ; Γ M : T — where Θ is a subtyping context, Γ is a sequence of typing assumptions x1 : S1 , . . . , xn : Sn and T is a type such that |Θ| S1 , . . . , Sn , T — are derived according to the rules in Table 2 extended with constants for arithmetic, conditional branching (If0S : nat → [S → [S → [S]]]) and reference declaration (newS : var[S]). The operational semantics of L≤ is based on that given for L in [3]. We extend the language with an unbounded set of constants a, b, . . . representing location names, and define the values of this extended language by the grammar: V ::= n | a | a.set | {l1 = M1 , . . . , ln = Mn } | λx.M | ΛX.M An environment E is a pair of a set of location names, and a partial function from this to the set of values. The evaluation relation ⇓ between pairs (M, S) of a closed term and environment and (V, S  ) of a value and environment is defined in Table 3. Table 3. Operational semantics for L≤

5.1

Denotational Semantics

The denotational semantics of L≤ extends the games interpretation of references [3] with our model of bounded quantification. We define the denotations of types and the subtyping relation first. Following [2, 3], we will interpret method types as indexed families of games. That is, we interpret types with n free variables in the category Fam(G(n))

66

J. Laird

(the coproduct completion of G(n)) in which objects are set-indexed families of objects of G(n), and morphisms from {Ai | i ∈ I} to {Bj | j ∈ J} are pairs (f : I → J, {σi : Ai → Af (i) | i ∈ I} of a reindexing function and an indexed family of morphisms from G(n)). Fam(G(n)) is Cartesian closed and has (small) coproducts [2]. Defining Σ{Ai | i ∈ I} = ∀n+1 (Πi∈I Jn+1 (Ai ) ⇒ •n+1 ) ⇒ •n+1 (as already observed in Sect. 2, this is concretely the same as the “lifted sum” construction used to interpret L in [3]) we have [15]: Proposition 8. Inclusion of G L (n) in Fam(G(n)) is right adjoint to Σ . This yields a strong monad on Fam(G(n)}. Object types are interpreted as objects of G(n) and method types as objects of Fam(G(n)), as follows: [[nat]]n = {1 | i ∈ N} [[[l1 : S1 , . . . , lm : Sm ]]]n = Πi≤m Σ[[Si ]]n [[Xi ]]n = •i

[[unit]] = {1} [[S → O]]n = [[S]]n ⇒ {[[O]]n } [[O]] [[∀Xn+1 ≤ O.P ]]n = ∀n [[P ]]n+1

Subtyping contexts X1 ≤ O1 , . . . , Xn ≤ On are interpreted as the corresponding bounding sequences [[O1 ]]0 , . . . , [[On ]]n−1 , and subtyping judgments Θ T ≤ T  L are interpreted as linear coercing morphisms from T to T  in Fam(G[[Θ]] ). The  key point here is that while Θ T ≤ T may have multiple derivations, they all correspond to the same morphism. Table 4. Denotational semantics of subyping judgments

Proposition 9. For any derivable subtyping judgment Θ T ≤ T  , there is a unique morphism [[Θ T ≤ T  ]] satisfying the rules in Table 4. Proof. We define a notion of canonical derivation for subtyping judgments by replacing the transitivity and reflexivity rules with: Θ X≤X

Θ,X≤O,Θ O≤O Θ,X≤O X≤O

(This is essentially Curien and Ghelli’s deterministic system for deriving subtyping judgments for F≤ [9].) We then prove that every derivation is denotationally equivalent to a canonical derivation. 5.2

Semantics of Terms

Terms in context are interpreted as morphisms in Fam(G[[Θ]] ) — terms of method type Θ; Γ M : S denote morphisms from [[Γ ]] to Σ[[S]], and terms Θ; Γ M : O

Game Semantics for Bounded Polymorphism

67

of object type denote morphisms from [[Γ ]] to [[O]]. The operations of L are interpreted as in the computational λ-calculus, and the constant new as the reference cell strategy defined in [3], while subsumption, universal quantification and instantiation are interpreted as follows: [[Θ; Γ M : O ]] = [[Θ; Γ M : O]]; [[Θ O ≤ O ]] [[Θ; Γ M : ∀X ≤ O.P ]] = ∀[[O]]Θ [[Θ, X ≤ O; Γ M : P ]] [[Θ; Γ M {O } : P [O /X]]] = [[Θ; Γ M : ∀X ≤ O.P ]]{[[Θ O ≤ O]], [[O ]]|Θ| } As in the semantics of subtyping judgments, these rules show how to interpret each derivation of a typing judgment as a morphism. We need to show that any derivation for a given term in context yields the same denotation. Proposition 10. For any derivable typing judgment Θ; Γ M : T , there is a unique morphism [[Θ; Γ M : T ]]. Proof. This follows the proof in [5]. Use of the subsumption rule generates multiple derivations of the same typing judgment, which are shown to be equivalent by extending the language with a constant convert : ∀X.[∀Y ≤ X.Y → X] (denoting the corresponding coercion), and replacing all uses of the subsumption rule with explicit coercions using convert. Using the (di)naturality properties of convert, we show that any two terms which correspond to the same term of L≤ obtained by erasing all occurrences of convert have the same denotation. We prove soundness of the operational rules using the properties of the computational λ-calculus, the equations relating the cell strategy to declaration, assignment and derefering in L [3] together with the properties of our semantics of bounded quantification (i.e. β-reduction for type-instantiation, which follows from Proposition 7). Proposition 11. If M, ⇓ V ; S then [[M ]] = ⊥. We prove computational adequacy using the approach described in [15]: defining an approximating semantics in which each cell can be accessed a bounded number of times, for which adequacy follows from the soundness of the operational rules by induction; this implies adequacy for the unbounded semantics by continuity. Proposition 12 (Computational Adequacy). [[M ]] = ⊥ implies M ⇓. Let T be the observational preorder at on closed terms of (closed) type T induced by the operational semantics — i.e. M T N if and only if for all contexts C[ : T ], C[M ] ⇓ implies C[N ] ⇓. (Note that if S ≤ T , terms may be observationally equivalent at type T but not at type S.) By a standard argument from adequacy: Corollary 1. If [[M : T ]] ⊆ [[N : T ]] then M T N .

68

6

J. Laird

Full Abstraction

We now consider how closely our model reflects the observational preorder and equivalence. First, we give a full abstraction result for a type-restricted fragment of the language. Define the concretely bounded types by the following grammar: S :: B | S → O | ∀X ≤ [l1 : S1 , . . . , ln : Sn ].O

O ::= X | [l1 : S, . . . , ln : S]

That is, we require that quantification bounds are not type variables. This is a significant constraint — it prevents the inheritance between variable types from being represented directly as a subtyping assumption — but leaves an expressive typing system. Proof that every finite strategy over a concretely bounded type is definable as a term closely follows the decomposition argument given in [15] for unbounded polymorphism (System F) with general references. The reason for the restriction on types is that we can eliminate negatively occuring concretely bounded quantifiers by instantiating them with a type which extends their bound with a single method of type unit — i.e. given a strategy σ : [[∀X ≤ [l1 : T1 , . . . , ln : Tn ].O]] → [[T ]], we can find a strategy σ  : [[[O[l1 : T1 , . . . , ln : Tn , l : unit]/X]]] → [[T ]] such that if σ  is the denotation of a term x : [O[[l1 : T1 , . . . , ln : Tn , l : unit]/X]] M : T , then σ is the denotation of y : ∀X ≤ [l1 : T1 , . . . , ln : Tn ] M [y{[l1 : T1 , . . . , ln : Tn , l : unit]/x]. Using this property, in conjunction with the decomposition argument given in [15], we may show that: Proposition 13. For any concretely bounded type T , context Γ and subtyping context Θ, each finite strategy σ : [[Γ ]]Θ → [[T ]]Θ is the denotation of a term Θ; Γ M : T . This is sufficient to establish full abstraction at concretely bounded types, following the argument from finite definability in [3]. For any strategy σ, let comp(σ) be the complete plays (sequences with no unanswered questions) of σ. Theorem 1. For any terms M, N : T , where T is a concretely bounded type, M T N if and only if comp([[M ]]) ⊆ comp([[N ]]). Finally, we give a counterexample showing that full abstraction does not hold at all types. Consider the strategy consisting of prefixes of the sequence: •1 ∀

1 (∀2 .Σ

(•2

PQ OA OA



•2 ))



Σ1) OQ

PQ PA

This strategy is not the denotation of any term X ≤ ; x : S M : [unit], where S(X) = ∀Y ≤ X.[Y → Y ]. Informally, if M calls x then it must instantiate it with a subtype of X — in the absence of further subtyping assumptions this

Game Semantics for Bounded Polymorphism

69

must be X itself, giving an object of type [X → X]. But M cannot use such a method, as X is unbounded and does not occur anywhere else among its types. Formally, we observe that: Lemma 3. For any term x : S M : O, where X does not occur in Γ, O [[Γ, x : S M : O]] = [[Γ, x : S M [ΛY.[⊥]/x] : O]] From this we may derive a counterexample to full abstraction. Let T = ∀X ≤ .[(S → [unit]) → [unit]] and consider the terms M = λf T .f {}ΛY.[⊥] and N = λf T .f {}ΛY.[λy Y .(y ())] of type T . Evidently, M T N , moreover the inclusion of [[M ]] in [[N ]] in the games model at this type is strict — they are separated by playing against the strategy defined above. Proposition 14. M1 and M2 are observationally equivalent at T → [unit]. Proof. It is sufficient to show that for any term V = ΛX.λx.M : T , V {} ΛY.λy Y .y is equivalent to V {}ΛY.λy Y .⊥. This follows from Lemma 3, since M is (denotationally and thus observationally) equivalent to M [ΛY.[⊥]/x]. Arguably, this example shows up an expressive deficit in L≤ (and similar type theories for bounded quantification such as F≤ ): it does not allow the extension of a variable type X with further methods to create a new subtype of X — in the absence of further assumptions, the only subtype of X is X itself, even though X represents an object which could be extended with new fields.

7

Further Directions

Avenues for further research include: – Our model for bounded quantification has some quite general aspects, being based on a model of unbounded quantification in which morphisms correspond to composable dinatural transformations. A systematic account of this construction, and its coherence properties remains to be given. – L≤ could be seen as a more general system with polymorphism at value (method) types, extending the unbounded case [14]. Subtyping for such types is based on inclusion of their values, sucggesting that quantifiecation is bounded below (e.g. by the empty type). – Extension of the full abstraction result to all types, by allowing the extension of a variable type with further methods. This would appear to require the capacity to declare new method labels, suggesting a role for nominal games. – An open problem is to give a directly defined subtyping relation on games which captures all subtyping judgments of L≤ (or F≤ itself). This appears feasible, notwithstanding the undecidability of subtyping in F≤ (which extends readily to L≤ ), since such a relation need not precisely coincide with subtyping in F≤ — for example, the types  and S ⇒  are syntactically distinct, but semantically equivalent. Acknowledgements. Research supported by EPSRC grant EP/K037633/1.

70

J. Laird

References 1. Abramsky, S., Jagadeesan, R.: A game semantics for generic polymorphism. Ann. Pure Appl. Logic 133(1), 3–37 (2004) 2. Abramsky, S., McCusker, G.: Call-by-value games. In: Nielsen, M., Thomas, W. (eds.) CSL 1997. LNCS, vol. 1414, pp. 1–17. Springer, Heidelberg (1998) 3. Abramsky, S., Honda, K., McCusker, G.: A fully abstract games semantics for general references. In: Proceedings of the 13th Annual Symposium on Logic in Computer Science, LICS 1998, IEEE Press (1998) 4. Bainbridge, E.S., Freyd, P.J., Scedrov, A., Scott, P.J.: Functorial polymorphism. Theor. Comput. Sci. 70(1), 35–64 (1990) 5. Bruce, K., Longo, G.: A modest model of records, inheritance and bounded quantification. Inf. Comput. 87(1/2), 196–240 (1990) 6. Cardelli, L., Wegner, P.: On understanding types, data abstraction and polymorphism. Comput. Surv. 17(4), 471–522 (1985) 7. Cardelli, L., Mitchell, J.C., Martini, S., Scedrov, A.: An extension of System F with subtyping. Inf. Comput. 109(1–2), 4–56 (1994) 8. Chroboczek, J.: Game semantics and subtyping. In: Proceedings of the Fifteenth Annual Symposium on Logic in Computer Science, pp. 192–203. IEEE press (2000) 9. Curien, P.-L., Ghelli, G.: Coherence of subsumption, minimum typing and typechecking in F. Math. Struct. Comput. Sci. 2(1), 55–91 (1992) 10. de Lataillade, J.: Dinatural terms in System F. In: Proceedings of the 24th Annual Symposium on Logic in Computer Science, LICS 2009, IEEE Press (2009) 11. Girard, J.-Y.: Linear logic. Theor. Comput. Sci. 50, 1–102 (1987) 12. Hughes, D.: Games and definability for System F. In: Proceedings of the Twelfth International Syposium on Logic in Computer Science, LICS 1997, IEEE Computer Society Press (1997) 13. Laird, J.: Game semantics for a polymorphic programming language. In: Proceedings of LICS 2010, IEEE Press (2010) 14. Laird, J.: Game semantics for call-by-value polymorphism. In: Abramsky, S., Gavoille, C., Kirchner, C., Meyer auf der Heide, F., Spirakis, P.G. (eds.) ICALP 2010. LNCS, vol. 6199, pp. 187–198. Springer, Heidelberg (2010) 15. Laird, J.: Game semantics for a polymorphic programming language. J. ACM 60(4) (2013) 16. McCusker, G.: Games and full abstraction for a functional metalanguage with recursive types. Ph.D. thesis, Imperial College London, Published by Cambridge University Press (1996) 17. Reynolds, J.C.: Towards a theory of type structure. In: Robinet, B. (ed.) Programming Symposium. Lecture Notes in Computer Science, vol. 19, pp. 408–425. Springer, Heidelberg (1974)

Recursion and Fixed-Points

Join Inverse Categories as Models of Reversible Recursion Holger Bock Axelsen and Robin Kaarsgaard(B) DIKU, Department of Computer Science, University of Copenhagen, Copenhagen, Denmark {funkstar,robin}@di.ku.dk Abstract. Recently, a number of reversible functional programming languages have been proposed. Common to several of these is the assumption of totality, a property that is not necessarily desirable, and certainly not required in order to guarantee reversibility. In a categorical setting, however, faithfully capturing partiality requires handling it as additional structure. Recently, Giles studied inverse categories as a model of partial reversible (functional) programming. In this paper, we show how additionally assuming the existence of countable joins on such inverse categories leads to a number of properties that are desirable when modelling reversible functional programming, notably morphism schemes for reversible recursion, a †-trace, and algebraic ω-compactness. This gives a categorical account of reversible recursion, and, for the latter, provides an answer to the problem posed by Giles regarding the formulation of recursive data types at the inverse category level.

1

Introduction

Reversible computing, that is, the study of computations that exhibit both forward and backward determinism, originally grew out of the thermodynamics of computation. Landauer’s principle states that computations performed by some physical system (thermodynamically) dissipate heat when information is erased, but that no dissipation is entailed by information-preserving computations [28]. This has motivated a long study of diverse reversible computation models, such as logic circuits [15], Turing machines [4, 6], and many forms of restricted automata models [27, 31]. Reversibility concepts are important in quantum computing, but are increasingly seen to be of interest in other areas as well, including high-performance computing [33], process calculi [13], and even robotics [34,35]. In this paper we concern ourselves with the categorical underpinnings of functional reversible programming languages. At the programming language level, reversible languages exhibit interesting program properties, such as easy program inversion [40]. Now, most reversible languages are stateful, giving them a fairly straightforward semantics interpretation. While functional programs are usually easier to reason about at the meta-level, they do not have the concept of state that imperative languages do, making their semantics interesting objects of study. c Springer-Verlag Berlin Heidelberg 2016  B. Jacobs and C. L¨ oding (Eds.): FOSSACS 2016, LNCS 9634, pp. 73–90, 2016. DOI: 10.1007/978-3-662-49630-5 5

74

H.B. Axelsen and R. Kaarsgaard

Further, many reversible functional programming languages (such as Theseus [25] and the Π-family of combinator calculi [7]) come equipped with a tacit assumption of totality, a property that is neither required [4] nor necessarily desirable as far as guaranteeing reversibility is concerned. Shedding ourselves of the “tyranny of totality,” however, requires us to handle partiality explicitly as additional categorical structure. One approach which does precisely that is inverse categories, as studied by Cockett & Lack [9] as a specialization of restriction categories, which have recently been suggested and developed by Giles [16] as models of reversible (functional) programming. In this paper, we will argue that assuming ever slightly more structure on these inverse categories, namely the presence of countable joins of parallel morphisms [17], gives rise to a number of additional properties useful for modelling reversible functional programming, notably two different notions of reversible recursion, and an account of recursive data types (via algebraic ω-compactness with respect to structure-preserving functors), which are not present in the general case. This is done by adopting two different, but complementary, views on inverse categories with countable joins as enriched categories – as CPO-categories, and as (specifically ΣMon-enriched) strong unique decomposition categories [18, 23]. Overview. The necessary background on restriction and inverse categories is presented in Sect. 2. In Sect. 3 we show that inverse categories with countable joins are CPO-enriched, which allows us to demonstrate the existence of (reversible!) fixed points of both morphism schemes and structure-preserving functors. In Sect. 4 we show that inverse categories with countable joins and a join-preserving disjointness tensor are unique decomposition categories equipped with a uniform †-trace. Sect. 5 gives conclusions and directions for future work.

2

Background

This section gives an introduction to restriction and inverse categories (with joins), dagger categories, and categories of partial maps as they will be used in the remainder of this text. Unless otherwise stated, the material described in this section can be found in standard texts on restriction and inverse category theory (e.g., Cockett & Lack [9–11], Giles [16], Guo [17]). We begin by recalling the definition of restriction structures and restriction categories. Definition 1 (Cockett & Lack, 2002). A restriction structure on a category consists of an operator (·) on morphisms mapping each morphism f : A → B to a morphism f : A → A (the restriction idempotent of f ) such that (i) (ii) (iii) (iv)

f ◦ f = f for all morphisms f : A → B, f ◦ g = g ◦ f whenever dom(f ) = dom(g), f ◦ g = f ◦ g whenever dom(f ) = dom(g), and h ◦ f = h ◦ f ◦ f whenever cod(f ) = dom(h).

A category with a restriction structure is called a restriction category.

Join Inverse Categories as Models of Reversible Recursion

75

As a trivial example, any category can be equipped with a restriction structure given by setting f = 1A for every morphism f : A → B. However, there are also many useful and nontrivial examples of restriction categories (see, e.g., Cockett & Lack [9, Sect. 2.1.3]), the canonical one being the category Pfn of sets and partial functions. In this category, the restriction idempotent f : A → A for a partial function f : A → B is given by the partial identity function  x if f is defined at x, f (x) = undefined otherwise. Since we take restrictions as additional structure, we naturally want a notion of functors that preserve this structure. Definition 2. A functor F : C → D between restriction categories C and D is a restriction functor if F (f ) = F (f ) for all morphisms f of C . A morphism f : A → B of a restriction category is said to be total if f = 1A . Given a restriction category C , we can form the category Total(C ), consisting of all objects and only the total morphisms of C , which embeds in C via a faithful restriction functor. Further, a restriction category in which every restriction idempotent splits is called a split restriction category, and, by way of the Karoubi envelope, every restriction category C can be embedded in a split restriction category Split(C ) via a fully faithful restriction functor. Restriction categories with restriction functors form a category, rCat. A useful feature of restriction categories, and one we will exploit throughout this article, is that hom-sets can be equipped with a partial order (often called the natural partial order), defined as follows: Proposition 1. In a restriction category C , every hom-set HomC (A, B) can be equipped with the structure of a partial order where we let f ≤ g iff g ◦ f = f . Further, every restriction functor F is locally monotone on this order, in the sense that f ≤ g implies F (f ) ≤ F (g). In Pfn, this corresponds to the usual partial order on partial functions: For f, g : A → B, f ≤ g if, for all x ∈ A, f is defined at x implies that g is defined at x and f (x) = g(x). A natural question to ask is when this partial order has a least element: A sufficient condition for this is when the restriction category has a restriction zero. Definition 3. A restriction category C has a restriction zero object 0 iff for all objects A and B, there exists a unique morphism 0A,B : A → B that factors through 0 and satisfies 0A,A = 0A,A . If such a restriction zero object exists, it is unique up to (total) isomorphism. When a given restriction category has such a restriction zero, the zero map 0A,B : A → B is precisely the least element of HomC (A, B). Moving on to inverse categories, in order to define these we first need the notion of a partial isomorphism:

76

H.B. Axelsen and R. Kaarsgaard

Definition 4. In a restriction category C , we say that a morphism f : A → B is a partial isomorphism if there exists a unique morphism f ◦ : B → A of C (the partial inverse of f ) such that f ◦ ◦ f = f and f ◦ f ◦ = f ◦ . Definition 5. A restriction category C is said to be an inverse category if all morphisms of C are partial isomorphisms. In this manner, if we accept an intuition of restriction categories as “categories with partiality,” inverse categories are “groupoids with partiality” – and, indeed, the category PInj of sets and partial injective functions is the canonical example of an inverse category. In fact, the Wagner-Preston representation theorem (see, e.g., Lawson [29]) for inverse monoids can be extended to show that every locally small inverse category can be faithfully embedded in PInj (see Heunen [21] for the general case, or Cockett & Lack [9] for the special case of small categories). The analogy with groupoids goes even further; similar to how we can construct a groupoid Core(C ) by taking only the isomorphisms of C , every restriction category C has a subcategory Inv(C ) that is an inverse category, and has as morphisms only the partial isomorphisms of C . Inverse categories with restriction functors form a category, invCat. More generally, inverse categories are dagger categories (sometimes also called categories with involution): Definition 6. A category C is said to be a dagger category if it is equipped with a contravariant endofunctor (−)† : C → C op such that 1†A = 1A for all objects A, and f †† = f for all morphisms f . Proposition 2. Every inverse category C is a dagger category with the dagger functor given by A† = A on objects, and f † = f ◦ on morphisms. As is conventional, we will call f † the adjoint of f , and say that f is selfadjoint if f = f † , and unitary if f † = f −1 . In inverse categories, unitary morphisms thus correspond precisely to (total) isomorphisms. For the remainder of this text, we will use this induced dagger-structure when refering to the partial inverse of a morphism (and write, e.g., f † rather than f ◦ ). A useful feature of this definition of inverse categories is that we do not need an additional notion of an “inverse functor” as a functor that preserves partial inverses; restriction functors suffice. Definition 7. A functor F : C → D between dagger categories is a dagger preserving if F (f )† = F (f † ) for all morphisms f of C . Proposition 3. Every restriction functor F : C → D between inverse categories C and D is dagger preserving. That this holds can be seen from the fact that the property of being a partial isomorphism is defined purely in terms of composition and restriction idempotents, both of which are preserved by restriction functors.

Join Inverse Categories as Models of Reversible Recursion

2.1

77

Joins and Compatibility

Given that hom-sets of restriction (and, by extension, inverse) categories are partially ordered, one may wonder when this partial order has joins. It turns out, however, that it does not in the general case, and that only very simple restriction categories have joins for arbitrary parallel morphisms. However, we can define a meaningful notion of joins for parallel morphisms if this operation is not required to be total, but only be defined for compatible morphisms. For restriction categories, this compatibility relation is defined as follows: Definition 8. Parallel morphisms f, g : A → B of a restriction category C are said to be restriction compatible if g ◦ f = f ◦ g; if this is the case, we write f  g. By extension, a set S ⊆ HomC (A, B) is restriction compatible if all morphisms of S are pairwise restriction compatible. This compatibility relation can be extended to apply to inverse categories by requiring that morphisms be compatible in both directions: Definition 9. Parallel morphisms f, g : A → B of an inverse category C are said to be inverse compatible if f  g and f †  g † ; if this is the case, we write f  g . A set S ⊆ HomC (A, B) is inverse compatible if all morphisms of S are pairwise inverse compatible. The familiar reader will notice that this definition differs in its statement from Guo’s [17, p. 98], who defined f  g in an inverse category C if f  g holds in both C and C op (relying on the observation that inverse categories are simultaneously restriction categories and corestriction categories). To avoid working explicitly with corestriction categories, however, we will use this equivalent definition instead. We define restriction categories with (countable) joins as follows: Definition 10 (Guo, 2012). A restriction category C is a (countable) join restriction category if it has a restriction zero object, and satisfies that for all (countable) restriction  compatible subsets S of all hom sets HomC (A, B), there exists a morphism s∈S s such that   (i) s ≤ s∈S s for all s ∈ S, and s ≤ t for all s ∈ S implies s∈S s ≤ t;   (ii) s∈S s; s = s∈S  s = (f ◦ s) for all f : B → X; and (iii) f ◦ s∈S   s∈S  s ◦ g = (s ◦ g) for all g : Y → A. (iv) s∈S s∈S In addition, we say that a restriction functor that preserves all thus constructed joins is a join restriction functor. As a concrete example, Pfn has joins of all restriction compatible sets; here, f  g iff whenever f and g are both defined at some point x, f (x) = g(x), and the join of a set of restriction compatible partial functions F is given by ⎞ ⎛    f (x) if there exists an f  ∈ F such that f  is defined at x, ⎝ f ⎠ (x) = undefined otherwise. f ∈F

78

H.B. Axelsen and R. Kaarsgaard

Notice that the compatibility relation ensures precisely that the result is a partial function. This, finally, allows us to define join inverse categories by narrowing the definition above to only require the existence of joins of inverse compatible (sets of) morphisms: Definition 11. An inverse category C is a (countable) join inverse category if it has a restriction zero object, and Definition 10 is satisfied for all (countable) inverse compatible subsets S of all HomC (A, B). Analogously to Pfn, the category PInj is a join inverse category with joins given precisely as in Pfn, since the additional requirement that f †  g † ensures that the resulting partial function is injective. 2.2

Categories of Partial Maps

Categories of partial maps provide a synthetic approach to partiality in a categorical setting, and was first introduced by Robinson and Rosolini [32] in 1988. To form a category of partial maps, we consider a stable system of monics: In a category C , a collection M of monics of C is said to be a stable system of monics if it contains all isomorphisms of C and is closed under composition and pullbacks (in the sense that the pullback m of an m : X → B in M along any f : A → B exists and m ∈ M). Given such a stable system of monics M in a category C , we can form the category of partial maps as follows: Proposition 4. Given a category C and a stable system of monics M of C , we form the category of partial maps Par(C , M) by choosing the objects to be the objects of C , and placing a morphism (m, f ) : A → B for every pair (m, f ) where m : A → A ∈ M and f : A → B is a morphism of C , as in

m

A

f

A

B

factored out by the equivalence relation · ∼ · in which (m, f ) ∼ (m , f  ) if there exists an isomorphism α : A → A such that m ◦ α = m and f  ◦ α = f . Composition of morphisms (m, f ) : A → B and (m , g) : B → C is given by (m ◦ m , g ◦ f  ) : A → C where m and f  arise from the pullback

m m A

A

f

A

B f

B

m



g C

where m ◦ m ∈ M precisely by M closed under composition and pullbacks. Categories of partial maps are prime examples of restriction categories; in fact, of split restriction categories. Further, it can be shown that every restriction category C embeds fully and faithfully and in a restriction preserving manner into a category of partial maps [9].

Join Inverse Categories as Models of Reversible Recursion

3

79

As CPO-categories

In the present section, we will show that inverse categories with countable joins are intrinsically CPO-enriched. This observation leads to two properties that are highly useful for modelling reversible functional programming, namely the existence of fixed points for both morphism schemes for recursion (that is, continuous endomorphisms on hom-objects) and for locally continuous functors. The former can be applied to model reversible recursive functions, and the latter to model recursive data types [3]. Further, we will show that the partial inverse of the fixed point of a morphism scheme for recursion can be computed as the fixed point of an adjoint morphism scheme, giving a style of recursion similar to the reversible functional programming language rfun [39]. Recall that a category C is CPO-enriched (or simply a CPO-category) if all HomC (A, B) are pointed ω-complete partial orders (i.e., they have a least element and satisfy that each ω-chain has a supremum), and composition is monotone, continuous and strict. To begin, we will need the lemma below. Lemma 1. Let C be an inverse category and f, g : A → B be parallel morphisms of C . If f ≤ g then f  g. This lemma allows us to show CPO-enrichment of join inverse categories: Theorem 1. Every inverse category C with countable joins is CPO-enriched. Proof. Let A, B be objects of C , and let {fi }i∈ω be an ω-chain in HomC (A, B) with respect to the canonical partial ordering. By Lemma 1, all fi and fj for i, j ∈ ω of this chain are inverse compatible, so the set F = {fi | i ∈ ω} is an inverse compatible subset of HomC (A, B). But then we can form the supremum of {fi }i∈ω by  sup{fi }i∈ω = f f ∈F

which is the supremum of this chain directly by definition of the join. Let f, g : A → B and F = {fi | i ∈ ω}. Monotony of compositions holds in all restriction categories, not just inverse categories with countable joins: Supposing f ≤ g then g ◦ f = f , and for h : B → X, h◦g◦h◦f =h◦g◦h◦g◦f =h◦g◦h◦g◦f =h◦g◦f =h◦f so h ◦ f ≤ h ◦ g in HomC (A, X); the argument is analogous for postcomposition. That composition is continuous follows directly by definition of joins, as we have   f= (h ◦ f ) = sup{h ◦ fi }i∈ω h ◦ sup{fi }i∈ω = h ◦ f ∈F

f ∈F

for all h : B → X, and analogously for postcomposition. That composition is strict follows by the fact that the zero map 0A,B : A → B is least in HomC (A, B), and that g ◦ 0A,B = 0A,X for all g : B → X by the universal mapping property of the zero object, and likewise for postcomposition.



80

H.B. Axelsen and R. Kaarsgaard

Recall that a functor F : C → D between CPO-categories is locally continuous iff each FA,B : HomC (A, B) → HomD (FA, FB) is monotone and continuous. Note that since all restriction functors preserve the partial order on hom-sets, and since suprema are defined in terms of joins, join restriction functors are in particular locally continuous. 3.1

Reversible Fixed Points of Morphism Schemes

In the following, let C be an inverse category with countable joins. We will use the term morphism scheme to refer to a monotone and continuous function f : HomC (A, B) → HomC (X, Y) – note that such schemes are morphisms of CPO and not of the inverse category C , so they are specifically not required to have inverses. Enrichment in CPO then has the following immediate corollary by Kleene’s fixed point theorem: f

→ HomC (A, B) Corollary 1. Every morphism scheme of the form HomC (A, B) − has a least fixed point fix f : A → B in C . Proof. Define fix f = sup{f n (0A,B )}; that this is the least fixed point follows by

Kleene’s fixed point theorem, as 0A,B is least in HomC (A, B). Morphism schemes on their own are useful for modelling parametrized reversible functions, i.e., functions that take other functions given at compile-time as parameters to produce new, first-order reversible functions. Since higher-order reversible functional programming is yet to be well-understood, parametrized functions (as implemented in, e.g., Theseus [25]) allow for a higher degree of abstraction and code reuse, as we know it from higher-order functional irreversible programming. With this in mind, recursive reversible functions can be seen as least fixed points of self-parametrized functions. Given that we can thus model reversible recursive functions via least fixed points of morphism schemes, a prudent question to ask is if the inverse of a least fixed point can be computed as the least fixed point of another morphism scheme. We will answer this in the affirmative, but to do this, we need to show that the induced dagger functor is locally continuous. Lemma 2. The canonical dagger functor † : C op → C is locally continuous. Proof. Let f, g : A → B. For monotony, suppose f ≤ g, i.e., g ◦ f = f . Then g† ◦ f † = g† ◦ f ◦ f † = g† ◦ g ◦ f ◦ f † = g ◦ f ◦ f † = g ◦ f ◦ f † = f ◦ f† = f† ◦ f ◦ f† = f† ◦ f† = f† so f † ≤ g † as well, as desired. For continuity, let {fi }i∈ω be an ω-chain in HomC (A, B), and let F = {fi | i ∈ ω} be the corresponding set for this chain.  †  Since f ≤ f ∈F f for each f ∈ F by Definition 11, we have f † ≤ f for f ∈F all f ∈ F by monotony of †, and so

Join Inverse Categories as Models of Reversible Recursion

sup{fi† }i∈ω

=



⎛ f†





f ∈F



81

⎞† f ⎠ = sup{fi }†i∈ω

f ∈F

 by Definition 11. In the other direction, we have f † ≤ f ∈F f † for all f ∈ F †  † by Definition 11, so by monotony of †, f = f †† ≤ f for all f ∈ F . f ∈F †   † But then f ∈F f ≤ by Definition 11, and so by monotony of †, we f ∈F f finally get ⎞†† ⎛ ⎞† ⎛    sup{fi }†i∈ω = ⎝ f⎠ ≤ ⎝ f †⎠ = f † = sup{fi† }i∈ω f ∈F

f ∈F

f ∈F



as desired.

With this lemma, we are able to show that the inverse of a least fixed point of a morphism scheme can be computed as the least fixed point of an adjoint morphism scheme: Theorem 2. Every f → HomC (A, B) has an adjoint mormorphism scheme of the form HomC (A, B) − f‡

phism scheme HomC (B, A) −→ HomC (B, A) such that (fix f )† = fix f‡ . Proof. Let ιA,B : HomC (A, B) → HomC (B, A) denote the family of functions defined by ιA,B (f ) = f † ; each of these are monotone and continuous by Lemma 2, and an isomorphism (with inverse ιB,A ) by f †† = f . Given a morphism scheme f : HomC (A, B) → HomC (A, B), we define f‡ = ιA,B ◦f ◦ιB,A – this is monotone and continuous since it is a (monotone and continuous) composition of monotone and continuous functions. But since f‡n = (ιA,B ◦ f ◦ ιB,A )n = ιA,B ◦ f n ◦ ιB,A by ιB,A an isomorphism with inverse ιA,B , and since 0†A,B = 0B,A by the universal mapping property of the zero object, we get fix f‡ = sup{f‡n (0B,A )} = sup{(ιA,B ◦ f n ◦ ιB,A )(0B,A )} = sup{f n (0†B,A )† } = sup{f n (0A,B )† } = sup{f n (0A,B )}† = (fix f )† as desired.



In modelling recursion in reversible functional programming, this theorem states precisely that the partial inverse of a recursive reversible functions is, itself, a recursive reversible function, and that it can be obtained by inverting the function body and replacing recursive calls with recursive calls to the thus constructed inverse: Coincidentally, this is precisely the inverse semantics of recursive reversible functions in rfun [39].

82

3.2

H.B. Axelsen and R. Kaarsgaard

Algebraic ω-compactness for Free!

A pleasant property of CPO-categories is that algebraic ω-compactness – the property that every locally continuous functor has a canonical fixed point – is relatively easy to check for, thanks to the fixed point theorem due to Ad´ amek [3] and Barr [5]: Theorem 3 (Ad´ amek & Barr). Let C be a CPO-category with an initial object. If C has colimits of ω-sequences of embeddings, then C is algebraically ω-compact over CPO. Canonical fixed points of functors are of particular interest in modelling functional programming, since they can be used to provide interpretations for recursive data types. In the following, we will couple this theorem with a joincompletion theorem for restriction categories to show that every inverse category can be faithfully embedded in an algebraically ω-compact inverse category with joins. That this succeeds rests on the following lemmas: Lemma 3. If an inverse category C embeds faithfully in a restriction category D, it also embeds faithfully in Inv(D). Proof. We notice that Inv : rCat → invCat is right adjoint to the forgetful functor U : invCat → rCat, with each component of the counit  : U ◦ Inv → 1rCat given by the faithful inclusion functor C : U (Inv(C )) → C (that this is an adjunction follows by an argument entirely analogous to Core : Cat → Grpd being right adjoint to U : Grpd → Cat). That faithful restriction functors are preserved follows readily since restriction functors preserve partial isomorphisms, and every restriction functor out of an inverse category factors through Invinclusion by this adjunction.

Lemma 4. The functor Inv : rCat → invCat takes join restriction categories to join inverse categories (and preserves join restriction functors). The latter of these lemmas was shown by Guo [17, Lemma 3.1.27]. To ease presentation of the completion theorem for join restriction categories due to Cockett and Guo [8, 17], we make the following notational shorthand: Convention 12. For a op restriction category C , let C denote the category of presheaves SetTotal(Split(C )) . Note in particular that C is cocomplete and all colimits are stable under pullback (since colimits in C are constructed object-wise in Set). Theorem 4 (Cockett & Guo). Every restriction category C can be faithfully  embedded in a join restriction category of the form Par(C , M gap ).  The stable system of monics M gap relates to the join-completion for restriction categories via M-gaps (see Cockett [8] or Guo [17, Sect. 3.2.2] for details). We can now show the algebraic ω-compactness theorem for restriction categories:

Join Inverse Categories as Models of Reversible Recursion

83

 Theorem 5. If C is a restriction category then Par(C , M gap ) is algebraically ω-compact for CPO; so for join restriction functors.  Proof. Let C be a restriction category. By Theorem 4, Par(C , M gap ) is join restriction category. By Ad´ amek & Barr’s fixed point theorem and the fact that join restriction categories are CPO-enriched (by Theorem 1) and have a restriction zero object (which is specifically initial) by definition, it suffices  to show that Par(C , M gap ) has colimits of ω-diagrams of embeddings. Let  D : ω → Par(C , M gap ) be such a diagram of embeddings. This corresponds to the diagram

m0

A

f0

m1

B

f1 D(2) . . .

D(1)

D(0)

in C . Since C is cocomplete, this diagram has a colimiting cocone α : D ⇒ colim D such that A B f0 f1 m0 m1

D(0)

D(2) . . .

D(1) αA αD(0)

αD(1) colim D

αB αD(2)

commutes. Further, since colimits in C are constructed object-wise in Set, this  colimit is stable under pullbacks, so composition in Par(C , M gap ), corresponding to pullbacks in C , commutes with this colimit. Thus, the family of morphisms 

{(mi , αD(i+1) ◦ fi )}i∈ω is a colimiting cocone for D in Par(C , M gap ). Finally, using this machinery, we can show how this theorem extends to inverse categories. Corollary 2. Every inverse category can be faithfully embedded in a join inverse category that is algebraically ω-compact for join restriction functors. Proof. Let C be an inverse category. Since U (C ) is the exact same category  viewed as a restriction category, U (C ) embeds faithfully in Par(C , M gap ), which is a join restriction category by Theorem 4, and algebraically ω-compact by Theorem 5. But then it follows by Lemma 3 that C embeds faithfully in  Inv(Par(C , M gap )), which is a join inverse category by Lemma 4, and is algebraically ω-compact for join restriction functors (which are specifically locally monotone and continuous) since fixed points of functors are (total) isomor

phisms, so preserved in Inv(Par(C , M gap )).

84

4

H.B. Axelsen and R. Kaarsgaard

As Unique Decomposition Categories

A complementary view on inverse categories with countable joins is as unique decomposition categories, a kind of category introduced by Haghverdi [18] equipped with a partial sum operation on Hom-sets via enrichment in the category of Σ-monoids (shown to be symmetric monoidal by Hoshino [23]). Unique decomposition categories (including Hoshino’s strong unique decomposition categories [23] which we will employ here) are specifically traced monoidal categories [26] if they satisfy certain conditions. This is desirable when modelling functional programming, as traces can be used to model notions of feedback [1] and recursion [19,20, 24]. Here, we will show that inverse categories with countable joins and a joinpreserving disjointness tensor (due to Giles [16]) are strong unique decomposition categories, and satisfy the conditions required to be equipped with a trace. We extend this result further to show that the trace is a †-trace [37], and thus has pleasant inversion properties (the trace in PInj is well studied, cf. [2, 18, 22]). This is particularly interesting given that the reversible programming language Theseus [25] and the combinator calculus Π 0 [7] both rely on a †-trace for reversible recursion. We begin with the definition of a Σ-monoid [18] (see also Manes & Benson [30] where these are described as positive partial monoids): Definition 13. A Σ-monoid (M, Σ) consists of a nonempty set M and a partial operator Σ defined on countable families in M (say that a family {xi }i∈I is

summable if i∈I xi is defined) such that (i) if {xi }i∈I is a countable family in M and {Ij }j∈J is a countable partitioning of I, then {xi }i∈I is summable iff all {xi }i∈Ij and i∈Ij xi are summable for all j ∈ J, and in this case   xi = xi , j∈J i∈Ij

i∈I

(ii) any family {xi }i∈I in M where I is singleton is summable with if I = {j}.

i∈I

xi = xj

The class of Σ-monoids with homomorphisms preserving partial sums forms a category, ΣMon. As such, a category C is enriched in ΣMon if all hom-sets of C are Σ-monoids, and composition distributes over partial addition. Lemma 5. Every inverse category with countable joins is ΣMon-enriched. Proof (Sketch). In an inverse category C , defining   si = s i∈I

s∈{si |i∈I}

for a countable family {si }i∈I of some hom-set HomC (A, B), summability coincides with inverse compatibility. That the axioms of Σ-monoids are satisfied follows straightforwardly by Definition 11.



Join Inverse Categories as Models of Reversible Recursion

85

Haghverdi defines unique decomposition categories in the following way: Definition 14 (Haghverdi). A unique decomposition category C is a symmetric monoidal category enriched in ΣMon such that for all finite index sets I and all j ∈ I, there are quasi-injections ιj : Xj → ⊕i∈I Xi and quasi-projections ρj : ⊕i∈I Xi → Xj satisfying (i) ρk ◦ ιj = 1Xk if j = k, and 0Xj ,Xk otherwise, and (ii) Σi∈I ιi ◦ ρi = 1⊕i∈I xXi . This definition is strengthened by Hoshino: Definition 15 (Hoshino). A strong unique decomposition category is a symmetric monoidal category enriched in ΣMon satisfying that the identity on the monoidal unit I is 0I,I , and 1X ⊕ 0Y,Y + 0X,X ⊕ 1Y = 1X⊕Y for all X and Y . An elementary result is that strong unique decomposition categories are unique decomposition categories, with their quasi injections and projections given by ι1 = (1A ⊕00,B )◦u−1 r : A → A⊕B and ρ1 = ur ◦(1A ⊕0B,0 ) : A⊕B → A, and analogously for ι2 and ρ2 (thus extending to any finite index set). As such, (strong) unique decomposition categories rely on a sum-like monoidal tensor – in the context of inverse categories, such a one can be found in Giles’ definition of a disjointness tensor [16, Definition 7.2.1]. Definition 16 (Giles). An inverse category C with a restriction zero object 0 is said to have a disjointness tensor if it is equipped with a symmetric monoidal restriction functor − ⊕ − : C × C → C such that (i) the restriction zero 0 is the tensor unit, and : A → A ⊕ B (ii) the morphisms given by 1 = (1A ⊕ 00,B ) ◦ u−1 r : A → B ⊕ A are jointly epic, and their partial and 2 = (00,B ⊕ 1A ) ◦ u−1 l inverses †1 : A ⊕ B → A and †2 : B ⊕ A → A are jointly monic, where ul : 0 ⊕ A → A and ur : A ⊕ 0 → A denote the left respectively the right unitor of the symmetric monoidal tensor. Though not required from this definition, since we are working exclusively with join inverse categories, we make the additional assumption that the disjointness tensor is a join restriction functor. Since Giles’ definition already demands that the zero object be the monoidal unit, and even defines i precisely like Hoshino’s definition of ιi (one can similarly see that †i = ρi ), we can show the following: Theorem 6. Every inverse category with countable joins and a join-preserving disjointness tensor is a strong unique decomposition category. Proof. By Lemma 5, any inverse category with countable joins (and a joinpreserving disjointness tensor) is enriched in ΣMon, so it suffices to show that the (specifically symmetric monoidal) disjointness tensor satisfies Definition 15. That 1I,I = 0I,I follows by 10,0 = 00,0 for the (restriction) zero 0, and 1X ⊕ 0Y,Y + 0X,X ⊕ 1Y = 1X⊕Y by the definition of partial sums as joins and the additional requirement that the disjointness tensor preserves joins.



86

H.B. Axelsen and R. Kaarsgaard

Due to the ΣMon-enrichment on unique decomposition categories, the trace can be constructed as a denumerable sum of morphisms, provided that morphisms of a certain form are always summable, cf. [18, Proposition 4.0.11] and [23, Corollary 5.4]: Theorem 7 (Haghverdi, Hoshino). Let C be a (strong) unique decomposition category such X, Y , and U and every f : X ⊕ U → Y ⊕ U ,

∞ that for every n ◦ f12 exists, where fij = ρj ◦ f ◦ ιi . Then C has a the sum f11 + n=0 f21 ◦ f22 uniform trace given by TrU X,Y (f ) = f11 +

∞ 

n f21 ◦ f22 ◦ f12 .

n=0

In a join inverse category, we say that parallel morphisms f, g : A → B are inverse disjoint if f ◦ g = 0A,A and f † ◦ g † = 0B,B . Lemma 6. In an inverse category, the following hold: (i) All inverse disjoint morphisms are inverse compatible, (ii) g  g  and f  f  implies g ◦ f  g  ◦ f  when dom(g) = cod(f ), and (iii) g ◦ f = g ◦ f ◦ f when dom(g) = cod(f ). This lemma allows us to show the existence of all trace sums: The core idea is to use part (ii) of this lemma until we get morphisms that are immediately disjoint by (iii), so inverse compatible by (i). Lemma 7. In a join inverse category with a disjointness tensor, all morphisms n ◦ f12 for any n ∈ N and some f : X ⊕ U → Y ⊕ U of the forms f11 or f21 ◦ f22 are pairwise inverse compatible. Recall that a †-category with a trace is said to have a †-trace (see, e.g., U † † Selinger [37]) if TrU X,Y (f ) = TrY,X (f ) for every morphism f : X ⊕ U → Y ⊕ U . Theorem 8. Every inverse category C with countable joins and a joinpreserving disjointness tensor has a uniform †-trace. Proof. By Theorem 6, C is a (strong) unique decomposition category, and by Lemma 7 it has all trace sums, so it follows that C has a uniform trace. To see that this is a †-trace, let f : X ⊕ U → Y ⊕ U be a morphism of C . We notice that † † † (fij )† = (ρj ◦ f ◦ ιi )† = ( †j ◦ f ◦ i )† = †i ◦ f † ◦ †† j = i ◦ f ◦ j = (f )ji

and so (f11 )† = (f † )11 and † n n n † (f21 ◦ f22 ◦ f12 )† = (f12 )† ◦ (f22 ) ◦ (f21 )† = (f † )21 ◦ (f22 ) ◦ (f † )12

which gives us

Join Inverse Categories as Models of Reversible Recursion

 TrU X,Y



(f ) =

f11 +



† f21 ◦

n f22

◦ f12



= (f11 ) ∨



f11 ∨

=

n∈ω †



f21 ◦

◦ f12

= (f11 )† ∨

n∈ω



= (f )11 ∨



n∈ω

†



f21 ◦

n f22

◦ f12

n∈ω

† n f22

87



n † (f12 )† ◦ (f22 ) ◦ (f21 )†

n∈ω †

(f )21 ◦

(f † )n22



◦ (f )12 =

† TrU Y,X (f )

  by definition of the partial sum as join (Lemma 5), and by ( f ∈F f )† = f ∈F f † by Lemma 2.



5

Conclusion

We have shown that inverse categories with countable joins carry with them a few key properties that are highly useful for modelling partial reversible functional programming. Notably, we have shown that any inverse category with countable joins is CPO-enriched – from this view, we gathered that morphism schemes have fixed points, and that the partial inverses of such fixed points can be computed as fixed points of adjoint morphism schemes. This gave us a model of recursion `a la rfun. Further, we were able to show that any inverse category can be embedded in an inverse category with joins, in which all join restriction functors have canonical fixed points. Finally, we showed that the presence of a join-preserving disjointness tensor on an inverse category with countable joins gives us a strong unique decomposition category, and in turn, a uniform †-trace: a model of recursion a` la Theseus and Π 0 . Restriction categories have recently been considered as enriched categories by Cockett & Garner [12], though their approach relied on enrichments based on weak double categories rather than monoidal categories, as it is otherwise usually done (including in this paper). Further, fixed points in categories with a notion of partiality have previously been considered, notably by Fiore [14] who also relied on order-enrichment, though his work was in categories of partial maps directly. Finally, Giles [16] has shown the construction of a trace in inverse categories recently, relying instead on the presence of countable disjoint sums rather than joins (whether or not this approach leads to a †-trace is unspecified). It should also be noted that the trace in the canonical inverse category PInj has seen study independent of unique decomposition and restriction categories, notably by Hines [22] and Abramsky, Haghverdi, and Scott [2]. As regards future work, since an inverse category with countable joins and a disjointness tensor is †-traced, it can be embedded in a †-compact closed category via the Int-construction [26,38]. It may be of interest to consider †compact closed categories generated in this manner, as we suspect these will be inverse categories as well (notably, Int(PInj) is [22]) – and could provide, e.g., an alternative treatment of projectors as restriction idempotents, and isometries as restriction monics (see also Selinger [36]).

88

H.B. Axelsen and R. Kaarsgaard

Additionally, while our focus in this article has been on inverse categories, we conjecture that many of these results can be generalized to restriction categories. Acknowledgments. We thank the anonymous reviewers for their useful comments. An abstract of this paper was presented at the 27th Nordic Workshop on Programming Theory held in Reykjav´ık, Iceland in October, 2015. The research was partly funded by the Danish Council for Independent Research | Natural Sciences under the Foundations of Reversible Computing project. We also acknowledge the support given by COST Action IC1405 Reversible computation.

References 1. Abramsky, S.: Retracing some paths in process algebra. In: Montanari, U., Sassone, V. (eds.) CONCUR 1996. LNCS, vol. 1119, pp. 1–17. Springer, Heidelberg (1996) 2. Abramsky, S., Haghverdi, E., Scott, P.: Geometry of interaction and linear combinatory algebras. Math. Struct. Comput. Sci. 12(5), 625–665 (2002) 3. Ad´ amek, J.: Recursive data types in algebraically ω-complete categories. Inf. Comput. 118, 181–190 (1995) 4. Axelsen, H.B., Gl¨ uck, R.: What do reversible programs compute? In: Hofmann, M. (ed.) FOSSACS 2011. LNCS, vol. 6604, pp. 42–56. Springer, Heidelberg (2011) 5. Barr, M.: Algebraically compact functors. J. Pure Appl. Algebra 82(3), 211–231 (1992) 6. Bennett, C.H.: Logical reversibility of computation. IBM J. Res. Dev. 17(6), 525– 532 (1973) 7. Bowman, W.J., James, R.P., Sabry, A.: Dagger traced symmetric monoidal categories and reversible programming. In: Proceedings of RC 2011, pp. 51–56. Ghent University (2011). http://www.cs.indiana.edu/∼sabry/papers/cat-rev.pdf 8. Cockett, J.R.B., Guo, X.: Join restriction categories and the importance of being adhesive (2007). http://www.mat.uc.pt/∼categ/ct2007/slides/cockett.pdf, presentation at Category Theory 2007 9. Cockett, J.R.B., Lack, S.: Restriction categories I: categories of partial maps. Theoret. Comput. Sci. 270(1–2), 223–259 (2002) 10. Cockett, J.R.B., Lack, S.: Restriction categories II: partial map classification. Theoret. Comput. Sci. 294(1–2), 61–102 (2003) 11. Cockett, J.R.B., Lack, S.: Restriction categories III: colimits, partial limits and extensivity. Math. Struct. Comput. Sci. 17(4), 775–817 (2007) 12. Cockett, R., Garner, R.: Restriction categories as enriched categories. Theoret. Comput. Sci. 523, 37–55 (2014) 13. Cristescu, I., Krivine, J., Varacca, D.: A compositional semantics for the reversible π-calculus. In: LICS 2013, pp. 388–397. IEEE Computer Society (2013) 14. Fiore, M.P.: Axiomatic Domain Theory in Categories of Partial Maps. Ph.D. thesis, University of Edinburgh (1994) 15. Fredkin, E., Toffoli, T.: Conservative logic. Int. J. Theor. Phys. 21(3–4), 219–253 (1982) 16. Giles, B.G.: An Investigation of some Theoretical Aspects of Reversible Computing. Ph.D. thesis, University of Calgary (2014) 17. Guo, X.: Products, Joins, Meets, and Ranges in Restriction Categories. Ph.D. thesis, University of Calgary (2012)

Join Inverse Categories as Models of Reversible Recursion

89

18. Haghverdi, E.: A Categorical Approach to Linear Logic, Geometry of Proofs and Full Completeness. Ph.D. thesis, Carlton University and University of Ottawa (2000) 19. Hasegawa, M.: Models of Sharing Graphs. Ph.D. thesis, University of Edinburgh (1997) 20. Hasegawa, M.: Recursion from cyclic sharing: traced monoidal categories and models of cyclic lambda calculi. In: de Groote, P., Hindley, J.R. (eds.) TLCA 1997. LNCS, vol. 1210, pp. 196–213. Springer, Heidelberg (1997) 21. Heunen, C.: On the functor 2 . In: Coecke, B., Ong, L., Panangaden, P. (eds.) Computation, Logic, Games and Quantum Foundations. LNCS, vol. 7860, pp. 107– 121. Springer, Heidelberg (2013) 22. Hines, P.M.: The Algebra of Self-Similarity and its Applications. Ph.D. thesis, University of Wales, Bangor (1998) 23. Hoshino, N.: A representation theorem for unique decomposition categories. In: Berger, U., Mislove, M. (eds.) MFPS XXVIII. Electronic Notes in Theoretical Computer Science, vol. 286, pp. 213–227. Elsevier (2012) 24. Hyland, M.: Abstract and concrete models for recursion. In: Grumberg, O., Nipkow, T., Pfaller, C. (eds.) Proceedings of the NATO Advanced Study Institute on Formal Logical Methods for System Security and Correctness, pp. 175–198. IOS Press (2008) 25. James, R.P., Sabry, A.: Theseus: a high level language for reversible computing, work-in-progress report at RC 2014 (2014). https://www.cs.indiana.edu/∼sabry/ papers/theseus.pdf 26. Joyal, A., Street, R., Verity, D.: Traced monoidal categories. Math. Proc. Camb. Philos. Soc. 119(3), 447–468 (1996) 27. Kutrib, M., Wendlandt, M.: Reversible limited automata. In: Durand-Lose, J., Nagy, B. (eds.) MCU 2015. LNCS, vol. 9288, pp. 113–128. Springer, Heidelberg (2015) 28. Landauer, R.: Irreversibility and heat generation in the computing process. IBM J. Res. Dev. 5(3), 183–191 (1961) 29. Lawson, M.V.: Inverse Semigroups: The Theory of Partial Symmetries. World Scientific, Singapore (1998) 30. Manes, E.G., Benson, D.B.: The inverse semigroup of a sum-ordered semiring. Semigroup Forum 31(1), 129–152 (1985) 31. Morita, K.: Two-way reversible multihead automata. Fundamenta Informaticae 110(1–4), 241–254 (2011) 32. Robinson, E., Rosolini, G.: Categories of partial maps. Inf. Comput. 79, 95–130 (1988) 33. Schordan, M., Jefferson, D., Barnes, P., Oppelstrup, T., Quinlan, D.: Reverse code generation for parallel discrete event simulation. In: Krivine, J., Stefani, J.B. (eds.) RC 2015. LNCS, vol. 9138, pp. 95–110. Springer, Heidelberg (2015) 34. Schultz, U.P., Bordignon, M., Støy, K.: Robust and reversible execution of selfreconfiguration sequences. Robotica 29(1), 35–57 (2011) 35. Schultz, U.P., Laursen, J.S., Ellekilde, L., Axelsen, H.B.: Towards a domain-specific language for reversible assembly sequences. In: Krivine, J., Stefani, J.B. (eds.) RC 2015. LNCS, vol. 9138, pp. 111–126. Springer, Heidelberg (2015) 36. Selinger, P.: Idempotents in dagger categories. In: Selinger, P. (ed.) QPL 2006. Electronic Notes in Theoretical Computer Science, vol. 210, pp. 107–122. Elsevier (2008)

90

H.B. Axelsen and R. Kaarsgaard

37. Selinger, P.: A survey of graphical languages for monoidal categories. In: Coecke, B. (ed.) New Structures for Physics. Lecture Notes in Physics, vol. 813, pp. 289–355. Springer, Heidelberg (2011) 38. Selinger, P.: Finite dimensional Hilbert spaces are complete for dagger compact closed categories. Logical Methods Comput. Sci. 8(3), 1–12 (2012) 39. Yokoyama, T., Axelsen, H.B., Gl¨ uck, R.: Towards a reversible functional language. In: De Vos, A., Wille, R. (eds.) RC 2011. LNCS, vol. 7165, pp. 14–29. Springer, Heidelberg (2012) 40. Yokoyama, T., Gl¨ uck, R.: A reversible programming language and its invertible selfinterpreter. In: Proceedings of the Partial Evaluation and Program Manipulation, pp. 144–153. ACM (2007)

A Coalgebraic View of Bar Recursion and Bar Induction Venanzio Capretta1 and Tarmo Uustalu2(B) 1 2

School of Computer Science, University of Nottingham, Nottingham, UK [email protected] Institute of Cybernetics, Tallinn University of Technology, Tallinn, Estonia [email protected]

Abstract. We reformulate the bar recursion and induction principles in terms of recursive and wellfounded coalgebras. Bar induction was originally proposed by Brouwer as an axiom to recover certain classically valid theorems in a constructive setting. It is a form of induction on nonwellfounded trees satisfying certain properties. Bar recursion, introduced later by Spector, is the corresponding function definition principle. We give a generalization of these principles, by introducing the notion of barred coalgebra: a process with a branching behaviour given by a functor, such that all possible computations terminate. Coalgebraic bar recursion is the statement that every barred coalgebra is recursive; a recursive coalgebra is one that allows definition of functions by a coalgebra-to-algebra morphism. It is a framework to characterize valid forms of recursion for terminating functional programs. One application of the principle is the tabulation of continuous functions: Ghani, Hancock and Pattinson defined a type of wellfounded trees that represent continuous functions on streams. Bar recursion allows us to prove that every stably continuous function can be tabulated to such a tree, where by stability we mean that the modulus of continuity is its own modulus. Coalgebraic bar induction states that every barred coalgebra is wellfounded; a wellfounded coalgebra is one that admits proof by induction.

1

Introduction

Bar induction is a reasoning principle formulated by L. E. J. Brouwer [9, 23]. He argued that it is justified in a constructive view of mathematics. Some classical theorems that are not otherwise provable in intuitionistic mathematics, follow from bar induction. Intuitively, it posits a link between termination and induction. It says that if processes of a certain class always terminate, then the class admits a form of wellfounded induction. Specifically, here is the original formulation of bar induction. Assume: – Q is a decidable predicate on lists of natural numbers and Q is a bar (for every stream σ there exists a natural n such that the list of the first n elements of σ satisfies Q); c Springer-Verlag Berlin Heidelberg 2016  B. Jacobs and C. L¨ oding (Eds.): FOSSACS 2016, LNCS 9634, pp. 91–106, 2016. DOI: 10.1007/978-3-662-49630-5 6

92

V. Capretta and T. Uustalu

– R is another predicate on lists of naturals, such that Q implies R (every list satisfying Q also satisfies R) and is inductive (for every list l, if every extension of l by one element satisfies R, then l also satisfies R); Then the empty list satisfies R. In this statement, streams must be interpreted as choice sequences, that is, infinite sequences of natural numbers not necessarily given by an effective rule. Kleene proved that bar induction implies that not all streams are computable [16]. Recently, Nakata, Bezem and one of us (Uustalu) [17] proved some interesting consequences of the principle on the relationships between some temporal operators (some of them mixed inductive/coinductive) on branching processes. Later, Spector [19] formulated the principle of bar recursion that allows definition of higher-order functionals (see also [10] and Chap. 5 of [4]); he exploited the principle to give a proof of consistency of mathematical analysis. Berardi, Bezem and Coquand [5] reformulated Spector’s principle and results; their version was termed modified bar recursion by Berger and Oliva [6], who also gave an analysis of the relation between the different variations. We offer a reformulation and generalization of the principle in terms of coalgebras. Coalgebras are useful to model branching computational processes and types of infinite data. In previous work [7, 8] we studied the notions of recursive and wellfounded coalgebras, useful to analyze recursive functions in total functional programming and induction proofs of their properties. Now we formulate a new notion of barred coalgebra, which characterizes processes whose computations always terminate, or data structures whose paths are all finite. We then state two principles as modern versions of Brouwer’s. Coalgebraic bar recursion states that every barred coalgebra is recursive. This allows us to define total recursive functions on a structure if we know that all its paths are finite. We give one application to continuous functions on streams. Ghani, Hancock and Pattinson [13] defined a type of inductive trees that represent continuous functions on streams by tabulating the outputs in their leaves. We restrict our attention to stably continuous functions, where stability means that the modulus of continuity is also continuous with itself as the modulus. We show that bar recursion implies that all stably continuous functions on streams can be tabulated. Coalgebraic bar induction states that every barred coalgebra is wellfounded. This allows us to reason about the process by a form of induction. The original form of bar induction is an instantiation of this version. We do not claim that the principles should be accepted always and in full generality. Instead, we view them as plausible assumptions encapsulating powerful function definition and reasoning devices that need to be justified depending on the exact setting where they are invoked. The paper is organized as follows. In Sect. 2, we introduce coalgebraic bar recursion. In Sect. 3, we apply it to tabulation of stably continuous stream functions. Coalgebraic bar induction is introduced and compared to Brouwer’s principle in Sect. 4.

A Coalgebraic View of Bar Recursion and Bar Induction

2

93

Bar Recursion

In previous work with Vene [7], we contributed to the study of recursive coalgebras, introduced by Osius [18] and further elaborated by Paul Taylor [20–22]. More recently, Ad´ amek, Milius and colleagues [2, 3] have made additional significant contributions. Definition 1. A coalgebra (A, α) of an endofunctor F on a category C is called recursive if for every algebra (C, γ) there exists a unique map f : A → C (a coalgebra-to-algebra morphism) making the following diagram commute: A

α

f

 Co

/ FA Ff

γ

 F C.

This is a useful notion in total functional programming: It guarantees that every structured recursive diagram (that is, a coalgebra-to-algebra morphism diagram) based on it is a definite description of a function. In the definition, unique existence of a mediating morphism is demanded upfront. No “more intrinsic” property of the coalgebra is invoked to guarantee unique constructibility of the solution. We are interested in coalgebras of functors that can be viewed as tree generators with a good notion of a path. A special class of set functors, called container functors [1] (they are closely related to polynomial functors [12]), give tree types that work for us. Definition 2. A container is given by a set S of shapes and an S-indexed family of positions for every shape. It defines a set functor S, P  by S, P  X = Σs : S. P s → X. We use containers to describe “branching types” of trees. A container (S, P ) says that there are S many “types” of branching nodes and that nodes of type s : S have P s many children. The sets of wellfounded and non-wellfounded trees with branching type (S, P ) are described as the inductive type μX. S, P  X (the initial algebra) and the coinductive type νX. S, P  X (the final coalgebra). For the first of these to have any elements at all, there must be at least one shape with no positions. We demand that inhabitedness of P s is decidable for all s : S. A coalgebra (A, α) is a process with a set of states A and a transition function α : A → S, P  A that generates a shape and new states in every position. Unfolding α takes a given initial state a : A to a non-wellfounded tree. This is the unique coalgebra morphism from (A, α) to the final coalgebra. Recursiveness of the coalgebra is equivalent to this tree unfolding being actually wellfounded for every initial state a : A, which in other words is to say that the unique coalgebra morphism to the final coalgebra factors through the initial

94

V. Capretta and T. Uustalu

algebra. (This was observed by Ad´ amek, L¨ ucke and Milius [2], who also called this condition the halting property.) We suggest a classically equivalent, but constructively weaker condition. This is that, for every initial state a : A, every path in the tree unfolding of α takes one from the root to a leaf in a finite number of steps.1 We need first some additional definitions to formalize what we mean by paths of an element of the coalgebra. For every state a, we define the set Pathα a of paths starting from it. At each stage, the path chooses a position through which to proceed. Notation: We define types and families by rules; sets of rules written with a single rule line are inductive definitions, sets of rules written with a double rule line are coinductive definitions. α a = (s, h) p : P s π : Pathα (h p) α a = (s, h) ¬P s end : Pathα a

p ≺ π : Pathα a

The first rule states that, if we reached a shape with no positions, then we are at a leaf of the tree and the path ends. The second rule states that we can construct a path by choosing a position in the present shape and continuing from the new state given by the transition in that position. This definition is coinductive, so the path may continue forever. Alternatively, we could define paths as functions from N; but this would exclude finite paths, which would have to be extended by iterating a distinguished element. The coinductive presentation is therefore more natural and avoids some coding. The status of general coinductive types in constructivism is unclear. However, paths and also tree unfoldings of coalgebras of containers can be coded with inductive and function types. So we are justified in using them. Next we inductively define finiteness of a path. A path π is considered finite (π ↓) if it reaches a positionless shape after a finite number of steps. The definition mirrors that of paths, but it is inductive. α a = (s, h) ¬P s end ↓

α a = (s, h)

p : P s π : Pathα (h p) p≺π ↓

π↓

Definition 3. A barred coalgebra is a coalgebra α whose all paths are finite: ∀a : A. ∀π : Pathα a. π ↓ . Note that the simpler condition that the coalgebra has no infinite paths, although classically equivalent, is constructively weaker than finiteness of all paths. Adopting this condition instead would make coalgebraic bar induction too strong; it would not follow from Brouwer’s formulation anymore. Assuming decidability of inhabitation of P s, the difference of the two conditions is exactly Markov’s principle. 1

In our constructive setting, it might in fact be more appropriate to take the name ‘halting’ for this condition that considers individual runs one by one rather than the whole entirety of evolutions of the process at once.

A Coalgebraic View of Bar Recursion and Bar Induction

95

We are now ready to formulate a coalgebraic version of the principle of bar recursion. It says that finiteness of all paths of a coalgebra implies that we can define functions by recursion. Classically this is provable, but constructively it is just a plausible extra axiom. Coalgebraic Bar Recursion: Every barred coalgebra is recursive. α barred ⇒ α recursive The converse implication does not need to be assumed: it is provable. Proposition 1. Every recursive coalgebra of a container functor is barred. α recursive ⇒ α barred Proof. Assume that a coalgebra α : A → S, P  A is recursive. Then we have the unique coalgebra-to-algebra morphism f : A → μX. S, P  X. For an element a : A, finiteness of all paths from a is proved by structural induction on f a.   We will not define Spector’s bar recursion here. (Instead we will discuss Brouwer’s bar induction in detail in Sect. 4.) But it corresponds to instances of coalgebraic bar recursion for S = 1 + 1 and P (inl ∗) = 0, P (inr ∗) = A, with A some fixed set. This means that S, P  X ∼ = 1 + (A → X). The typical case is A = N, but Spector also considered general A. This allowed him to interpret the general axiom of countable choice.

3

Continuous Functions on Streams

We illustrate the coalgebraic bar recursion principle by applying it to continuous functions on streams. A function on streams is continuous if it only uses a finite initial segment of its input stream to determine its result. Ghani, Hancock and Pattinson [13] defined an inductive type of trees for representing continuous functions on streams. When applied to a stream, a function can either immediately return a result, or read the next element of the stream and continue the computation. In the tree representation, an immediate return is modelled by a leaf containing the output value, a reading operation is modelled by a node that branches according to the input value. The statement that all continuous functions can be represented as trees in this manner is not provable constructively without additional assumptions. Ghani, Hancock and Pattinson give a proof of a negative version of this statement: If a function has no tree representation, then it cannot be continuous. We will show that, assuming coalgebraic bar recursion, the positive statement of representability becomes provable for what we call stably continuous functions. Stability is a natural condition, requiring that the modulus of continuity is also continuous, with itself as its modulus. Let SA be the type of streams (infinite sequences) of elements of type A. The equivalence relation =n , for n a natural number, identifies streams that coincide on the first n elements: σ1 =n σ2

if and only if

∀i < n. σ1 (i) = σ2 (i).

96

V. Capretta and T. Uustalu

Definition 4. A function f : SA → B from streams of elements of type A to results of type B is continuous if ∀σ : SA . ∃n : N. ∀σ  : SA . σ  =n σ → f σ  = f σ. So a function f is continuous on a stream σ if the value of f only depends on the first n elements of σ, for some n. It is continuous globally if it is continuous on every stream. This definition corresponds to SA being assigned the prodiscrete and B the discrete topology. This is what is appropriate for our purposes here. It can feel limited. For B = SA , for instance, not even the identity function is continuous, but then with the trees considered here it cannot be tabulated either.2 For other purposes, other choices can be appropriate. Brouwer’s continuity principle states that all functions of type SN → N are continuous. It seems justified by a computational view of functions as programs: a terminating program computes its result in a finite number of steps; during the computation it can only read a finite number of entries from the stream. This principle can be generalized to functions A → B where A and B are assigned their “native” topologies (whereby stream types must get product topologies). However, Escard´o recently discovered that the continuity principle is inconsistent in type theory [11], so the principle cannot be added safely to current type-theoretic foundations. The intuitive reason is that adding the principle adds new functions to the system and some of those are problematic. Various strands of research are investigating weaker and more refined versions of the principle that may be safe. For our goals, it is important that a continuous function has an explicit modulus of continuity, the mapping that gives the length of the initial segment of the stream needed for the computation, and that this modulus is stable, that is, it makes consistent choices for streams that have the same initial segment. Definition 5. A modulus of continuity for a function f : SA → B is a function mf : SA → N such that ∀σ, σ  : SA . σ  =mf σ σ → f σ  = f σ. The modulus is stable if ∀σ, σ  : SA . σ  =mf σ σ → mf σ  = mf σ. So a modulus of continuity is stable if and only if it is its own modulus of continuity. Stability is a reasonable assumption in the computational view of functions: – If, to compute f σ, we only need to read the first mf σ elements of σ; and – σ  coincides with σ on the first mf σ elements; 2

In a different work [14], Hancock, Pattinson and Ghani used a mixed inductive/coinductive type to tabulate stream processors, i.e., continuous functions f : SA → SB where both SA and SB are given the prodiscrete topology.

A Coalgebraic View of Bar Recursion and Bar Induction

97

– then we only need the first mf σ elements of σ  to compute f σ  (which is equal to f σ); – so it is reasonable to expect that mf σ  = mf σ. However, given a possibly non-stable modulus, it is not in general possible to construct another modulus which is stable. In the case that the original modulus is continuous, there is an algorithm to stabilize it. Given a non-stable but continuous modulus m, we construct a new stable modulus m ¯ for the same function. We do it by truncating every stream at an appropriate point, dictated by m, and filling in the rest of it with a fixed stream. If A is inhabited, so we know an element, we can repeat that element in a constant stream σ0 . For every σ and every index i : N, let σ|i be its truncation at position i, that is, the list [σ(0), . . . , σ(i − 1)]. Let ni = m (σ|i ++ σ0 ). (The notation + + denotes prepending a list to a stream and also to a path.) We now define the result of the new modulus m ¯ on σ: m ¯ σ = nk

where k = min{i | ni ≤ i}.

Continuity of m guarantees that m ¯ is well defined: k always exists. In fact we know that there is a j such that, if σ  =j σ, then m σ  = m σ. In particular, + σ0 where i = max(j, m σ), we have: if we choose σ  = σ|i + ni = m σ  = m σ ≤ i. So, since there is at least one i such that ni ≤ i, there is a minimal one k. In Ghani, Hancock and Pattinson’s work [13], a continuous function f : SA → B is represented by a wellfounded tree, an element of the inductive type SFA,B given by the rules b:B g : A → SFA,B write b : SFA,B read g : SFA,B The idea is that an element of SFA,B is a tabulation of all the values of a function as leaves in a tree whose finite paths have to be matched against the input stream. The application of a tabulation gives us a continuous function: apply : SFA,B → SA → B apply (write b) s = b apply (read g) (a ≺ s) = apply (g a) s The operator apply is defined by structural recursion on its tree argument. Is there an inverse transformation, that is, a tabulation operator assigning a tree to every continuous function, tabulate : (SA → B) → SFA,B ? Classically, we can prove that this function exists, but the proof is not constructive. Intuitively, the algorithm to obtain the tabulation should be the following. Let f : SA → B be continuous. Then we can define: tabulate f = write b if f “must be” constantly b tabulate f = read (λa.tabulate (λσ.f (a ≺ σ))) otherwise.

98

V. Capretta and T. Uustalu

It is of course undecidable whether f is constant. But if the function is continuous constructively, we also have a modulus for it and can check whether the modulus is 0 on some fixed stream (we assume A to be inhabited, so we can construct one). This a sufficient condition for constancy. However, it is not constructively provable that this algorithm generates a wellfounded tree. If the modulus is stable, we can prove that the paths of the recursive calls generated by the second equation for f are always finite. But this is not enough to conclude that the a priori non-wellfounded tabulation tree is wellfounded. Ghani, Hancock and Pattinson proved a negative version of the statement: they did not construct a tabulation operator, but proved that, if a function cannot be tabulated, then it is not continuous. With bar recursion, we can conclude the positive statement, using stable continuity. We provide two distinct proofs. The first proof is local: for every stably continuous function it constructs a barred coalgebra and uses it to generate the tabulation tree. The second proof is global: it constructs a single barred coalgebra on the set of all stably continuous functions and uses it to define a tabulation operation. 3.1

Individual Tabulations

First we show how to tabulate a single stably continuous function by associating an individual barred coalgebra to it. Let f : SA → B be a function with a stable modulus of continuity m : SA → N. We will assume that A is inhabited and has a distinguished element, which we can repeat in a stream σ0 . We construct a coalgebra on the functor F X = B + (A → X) with carrier the set of lists of elements of A: αf : List A → B + (A → ListA ) inl (f (l + + σ0 )) if m (l + + σ0 ) ≤ length l αf l = inr (λa. l + +[a]) otherwise. The coalgebra checks whether the list l is sufficient to determine the result given by f : if the modulus of continuity on a stream starting with l is at most the length of l, then we know that f will depend only on it. In this case, the coalgebra terminates with the value given by f . Else, it branches into new processes for all elements of A, lengthening the list by appending the element at the end. Observe that the functor F is a container with shapes B + 1, a shape for each possible leaf in B, and a single shape for the continuation. The shapes in B have no positions: the coalgebra terminates. The single continuation shape has positions for all possible input elements, so the set of positions is A. The paths in Pathα l are sequences of position choices, so in this case they are sequences of elements of A. Any such path π can be padded out to a stream of elements of A by appending σ0 to it: pad end = σ0 , pad (a ≺ π) = a ≺ pad π. Lemma 1. αf is a barred coalgebra.

A Coalgebraic View of Bar Recursion and Bar Induction

99

Proof. Given l : ListA and π : Pathα l, we prove that π ↓ by induction on + pad π. m σl,π − length l where σl,π = l + – If m σl,π − length l = 0, then m σl,π ≤ length l, so σl,π =m σl,π l ++ σ0 . By stability, m (l + + σ0 ) = m σl,π ≤ length l. Therefore αf l = inl (f (l ++ σ0 )) and π = end. In this case obviously the path is finite. – Otherwise, if m σl,π − length l > 0, then m σl,π > length l. It is impossible that m (l + + σ0 ) ≤ length l because, if it were, then also m σl,π = m (l ++ σ0 ) ≤ +[a]) and π = a ≺ π  for some a : A length l by stability. So αf l = inr (λa. l +  +[a]). and π : Pathα (l + Note that σl ++[a],π = σl,π , so m σl ++[a],π −length (l + +[a]) = m σπ −length l−1.   Therefore π  ↓ by the induction hypothesis and so π ↓. We can now invoke bar induction on this coalgebra to tabulate f . Theorem 1. The coalgebraic bar induction principle implies that there exists an element tabulatef : SFA,B such that apply (tabulatef ) σ = f σ for every σ : SA . Proof. If we apply the coalgebraic bar recursion principle to the statement of Lemma 1, we obtain that αf is a recursive coalgebra. Notice that SFA,B is the carrier of the initial algebra of the functor F : SFA,B = μX. B + (A → X). The copair of the two constructors [write, read] is the actual algebra. We can apply the defining property of recursive coalgebras to deduce that there exists a unique function tab making the following diagram commute. ListA

αf

/ B + (A → ListA )

tab

F tab



SFA,B o

[write,read]

 B + (A → SFA,B ).

We can now define the Ghani, Hancock and Pattinson representation of f by tabulatef : SFA,B tabulatef = tab []. Its correctness can be checked simply by observing that, in general, apply (tab l) σ = f (l ++ σ). 3.2

 

Global Tabulation

The second way to use coalgebraic bar recursion for tabulation is to define a coalgebra on the set of all stably continuous functions: F = {(f, m) | m : SA → N is a stable modulus of f : SA → B}.

100

V. Capretta and T. Uustalu

We use the same functor as for the individual coalgebras of the previous section: F X = B + (A → X). The algebra simply tests whether a function must be constant: if it has to be, it returns its value, otherwise it branches. It does so by checking that the modulus is 0 for σ0 ; it must be then be 0 for all streams by stability. α : F → B + (A → F) if m σ0 = 0 inl (f σ0 ) α (f, m) = inr (λa. (fa , ma ) otherwise where fa σ = f (a ≺ σ) and ma σ = m (a ≺ σ) − 1. In the branching case, the coalgebra reads an element of the input, a, and returns the function fa obtained by shifting f by a. The modulus ma of fa is one less than the modulus of f , as we do not count the prepended element a. Lemma 2. The coalgebra α is barred. Proof. Let π : Pathα (f, m) be a path of the coalgebra. We prove that π ↓ by induction on m σπ where σπ = pad π. – If m σπ = 0, then also m σ0 = 0 by stability. So α (f, m) = inl (f σ0 ) and π = end. – If m σπ > 0, then also m σ0 > 0 by stability. So α (f, m) = inr (λa. (fa , ma )). It must then be that π = a ≺ π  for some a : A and π  : Pathα (fa , ma ). If we then compute the modulus of the stream associated to the tail of the path, we obtain ma σπ = m (a ≺ σπ ) − 1 = m σπ − 1. Therefore π  ↓ by the induction hypothesis, so π ↓.

 

We can now invoke bar induction on this coalgebra to construct a global tabulation operator. Theorem 2. The coalgebraic bar induction principle implies that there exists a function tabulate : F → SFA,B such that, for every continuous function with stable modulus (f, m) : F and every stream σ : SA , we have apply (tabulate (f, m)) σ = f σ. Proof. Applying the coalgebraic bar induction principle to the result of Lemma 2, we obtain that (F, α) is a recursive coalgebra. In particular, the following diagram has a unique solution. F tabulate

α

F tabulate



SFA,B o

/ B + (A → F)

[write,read]

 B + (A → SFA,B ).

The operator tabulate maps every function to a correct tabulation tree for it. We prove this by induction on m σ0 :

A Coalgebraic View of Bar Recursion and Bar Induction

101

– If m σ0 = 0, then, by commutativity of the diagram, definition of α and continuity, apply (tabulate (f, m)) σ = apply ([write, read] (F tabulate (α (f, m)))) σ = apply ([write, read] (inl (f σ0 ))) σ = apply (write (f σ0 )) σ = f σ0 = f σ. – If m σ0 > 0, then, writing σ = a ≺ σ  , by commutativity of the diagram, the definition of α and the induction hypothesis, apply (tabulate (f, m)) σ = apply ([write, read] (F tabulate (α (f, m)))) σ = apply ([write, read] (inr (λa. tabulate (fa , ma )))) σ = apply (read (λa. tabulate (fa , ma ))) (a ≺ σ  ) = apply (tabulate (fa , ma ))) σ  = fa σ  = f (a ≺ σ  ) = f σ.

 

We finish this discussion of tabulation of continuous functions on streams by noting that we need bar recursion for tabulation because our notion of (stable) continuity of a function (which is the standard notion of continuity) is pathbased. It considers one stream (choice sequence) at a time and requires that reaching an answer takes a finite number of steps. Alternatively, we could define continuity in a way that considers the entirety of possible evolutions of the choice process at once. This would result in a stronger notion of continuity and then a continuous function could be tabulated without assumptions like bar recursion. Concretely, we could define continuity as a predicate on SA → B inductively as follows. ∀σ, σ  : SA . f σ = f σ  f continuous

∀a : A. (λσ. f (a ≺ σ)) continuous f continuous

Tabulation would be immediate by structural recursion on the proof of continuity (and not exciting at all).

4

Bar Induction

We now want to give a coinductive account of the traditional formulation of bar induction. For this purpose, the notion of wellfounded coalgebra will be useful. Intuitively, it is a coalgebra that admits proofs of properties of its elements by induction. This notion was introduced by Taylor [20–22], who proved that, under weak reasonable assumptions, recursiveness and wellfoundedness of a coalgebra are equivalent. (In a previous article [8], we gave a review of the topic and extended it to the dual case or corecursive vs. antifounded algebras.)

102

V. Capretta and T. Uustalu

Our formulation uses the next-time operator by Jacobs [15] on subobjects (subsets) of the carrier of a coalgebra. Let (A, α) be a coalgebra of a functor F that preserves pullbacks along monos on a category with pullbacks along monos. (These requirements are satisfied by container functors.) Definition 6. Let j : U → A be a subobject of the carrier of the coalgebra. The next-time subobject, ntα j : ntα U → A is defined by the following pullback. ntα  U _

α|j

ntα j

 A

/FU _ 

α

Fj

/FA

The idea of this definition is that, if U is a subset of A, then ntα U is the subset of A consisting of the elements that, after an α transition, fall into U . In particular, suppose F is a container functor, F = S, P . If a : A, then α a : F A has the form (s, h) with s : S and h : P s → A. We have that a ∈ ntα U if, for all p : P s, h p ∈ U . ntα U = {a : A | ∀p : P s. h p ∈ U where (s, h) = α a} Definition 7. The coalgebra (A, α) is wellfounded if, for every subobject j : U → A, if ntα U factors through U , then j is an isomorphism. In diagram form this says that ntα U Jr JJ JJ JJ ntα j JJ J%

A

/ wU Mm w ww w ww j w {w

=⇒

j is iso.

In simpler terms, the coalgebra is wellfounded, if ntα U ⊆ U implies A ⊆ U . Intuitively, the defining property of wellfounded coalgebras is a generalization of the familiar induction principle associated to initial algebras. We want to prove that all elements of A are in the subset U . And we do this by showing that, if all “components” of an element a : A (given by α a) are in U , so is a. We introduce the following principle. Coalgebraic Bar Induction: Every barred coalgebra is wellfounded. α barred ⇒ α wellfounded This principle says that, if all paths of a coalgebra are finite, then we can prove its properties by induction. The converse implication holds without extra assumptions. Proposition 2. Every wellfounded coalgebra is barred. α wellf ounded ⇒ α barred

A Coalgebraic View of Bar Recursion and Bar Induction

103

Proof. Assume that α is wellfounded and take U to be the subset of those elements of A whose all paths are finite: U = {a : A | ∀π : Pathα a. π ↓}. We then have that ntα U = {a : A | ∀p : P s. ∀π  : Pathα (h p). π  ↓ where (s, h) = α a}. Suppose a ∈ ntα U ; we want to prove that a ∈ U . So assume π : Pathα a; we will show that this path is finite. Suppose α a = (s, h). If π = end, as ¬P s, then we conclude immediately that π ↓. If π = p ≺ π  for some p : P s and π  : Pathα (h p), then we know by assumption that π  ↓ and consequently π ↓. We proved that ntα U ⊆ U ; therefore, by wellfoundedness of α, we have that A ⊆ U . So all paths are finite.   The traditional formulation of bar induction from Brouwer is about predicates on lists of natural numbers. Brouwer’s Bar Induction: – Let Q be a decidable predicate on lists of natural numbers. Assume that Q is a bar: ∀σ : SN . ∃n : N. Q (σ|n ). – Let R be a predicate on lists of naturals. Assume that Q implies R and R is inductive: ∀l : ListN . (∀p : N. R (l ++[p])) → R l. – Then R [] holds. We can assume that Q and R are suffix closed: ∀l : ListN . ∀p : N. Q l → Q (l + +[p]) and similarly for R. Indeed, if they are not, we can use their suffix ˆ ˆ l to hold if ∃n ≤ length l. Q (l|n ) and similarly for R. closures. We can define Q ˆ ˆ The new predicates Q and R still satisfy the assumptions of the principle and ˆ [] ↔ R []. Also, since Q is decidable, Q being a bar implies that it is also a R “tight” bar: ∀σ : SN . ∃n : N. (∀m : N. m < n → ¬Q (σ|m )) ∧ Q (σ|n ). Instead of N, one could consider other sets in Brouwer’s bar induction. The case of B is known as the fan “theorem” (Brouwer thought that fan theorem and bar induction were theorems, but both are just plausible axioms; the fan theorem is a positive version of weak K¨onig’s lemma). In principle one could replace N with any fixed set, but the question is how justified this is from the constructive point of view. N is special in that it is a set that can be traversed by an infinite process. The traditional form of bar induction follows from the coalgebraic version. Theorem 3. Coalgebraic bar induction implies Brouwer’s bar induction. Proof. Given Q and R satisfying the assumptions of Brouwer’s bar induction, we define a coalgebra structure of the functor F X = 1 + (N → X) on the set ListN . α : List N → 1 + (N → ListN ) inl ∗ if Q l αl = inr (λp. l + +[p]) otherwise.

104

V. Capretta and T. Uustalu

The function α is welldefined, since Q is decidable. From the assumption that Q is a bar, we conclude that α is barred. Indeed, let π : Pathα l for some list l; we must prove that π is finite. We consider the stream σ = l ++ pad π (using, e.g., 0 as the distinguished element of N for padding). Since Q is decidable and a bar, σ must have a shortest finite prefix σ|n satisfying Q. If n < length l, then σ|n is a proper prefix of l. This forces that π = end, since suffix-closedness of Q gives + end. Either way, π is finite. us that Q l. If n ≥ length l, then l + + π = σ|n + Thus we can apply coalgebraic bar induction and learn that α is wellfounded. Let U = {l : ListN | R l}. (In type theory we can use the dependent pair type Σl : ListN . R l.) We will prove that ntα U ⊆ U from where by wellfoundedness of α it will follow that ListN ⊆ U . By definition nt U = {l : ListN | α l ∈ 1 + (N → U )} We prove that if l ∈ nt U , then l ∈ U , by cases: – If Q l holds, then α l = inl ∗. Since Q implies R, we also have R l, that is, l ∈ U. – If Q l does not hold, then α l = inr (λp. l + +[p]) with l + +[p] ∈ U for every p : N. By the assumption that R is inductive, we can derive that l ∈ U . Therefore, since α is wellfounded, we can deduce that ListN ⊆ U , that is, ∀l : ListN . R l. In particular, R [], as desired.   From Brouwer’s bar induction, coalgebraic bar induction follows for the functor F X = 1 + (N → X). Theorem 4. Brouwer’s bar induction implies coalgebraic bar induction for F X = 1 + (N → X). Proof. Given any set A with a barred coalgebra structure α : A → 1 + (N → A) and a subset U of A satisfying the assumption ntα U ⊆ U of coalgebraic bar induction. We define a function  : A → ListN → 1 + A by a  [] = inr a

⎧ if a  l = inl ∗ ⎨ inl ∗ if a  l = inr a and α a = inl ∗ a  (l + +[p]) = inl ∗ ⎩ inr (f p) if a  l = inr a and α a = inr f For any a : A, we define Qa l to hold, if a  l = inl ∗, and Ra l to hold, if a  l = inl ∗ or a  l = inr a and a ∈ U . It is immediate that, for any a : A, Qa is decidable and ∀l : ListN . Qa l → Ra l. Moreover, as α is barred, Qa is also a bar. From ntα U ⊆ U , it also follows that Ra is inductive for all a : A. Hence, for any given a : A, we can apply Brouwer’s bar induction, and   conclude that Ra []. Since a  [] = inr a, this means that a ∈ U .

A Coalgebraic View of Bar Recursion and Bar Induction

105

Coalgebraic bar induction follows from Brouwer’s bar induction also for general container functors S, P  for which Σs : S. P s is isomorphic to a decidable subset of N. This is proved by “approximate” branchings of type S, P  with branchings of type F X = 1 + (N → X).

5

Conclusions

The explicit use of bar recursion and bar induction makes it possible to encapsulate non-constructive aspects of arguments about non-wellfounded trees and stream functions. In this paper, we have defined coalgebraic versions of bar recursion and bar induction. We find that this perspective has at least two benefits. First, we can speak of the ingredients involved in bar recursion and bar induction using the terminology of modern theoretical computer science, especially coalgebra. Second, we can avoid unnecessary coding when we want to deal with types of trees with multiple types of branching nodes, with different numbers of children. The notion of barred coalgebra neatly fills a useful position in the range of properties that a coalgebra may satisfy. In reasoning about the total correctness of programs, it is usual to give a proof of termination for all computations and then prove by induction that some invariant holds. This application of wellfounded induction is classically justified from termination, but is not constructively valid. Bar induction is the missing principle that provides the link. Since coalgebraic methods are becoming standard tools in programming and reasoning about correctness, we hope that the coalgebraic formulation of bar recursion will establish itself as a useful tool in functional programming and type theory. Directly applying the traditional forms of bar induction and bar recursion in concrete cases often requires a lot of encoding, namely making all nodes to be of the same type with the same number of children, which also makes all paths infinite, all this artificially; termination is then characterized by a predicate with the bar property. Our formulation is more liberal and requires no encoding: nodes of different branching degrees are allowed and termination is simply modelled by leaves. Acknowledgements. Capretta is grateful to the School of Computer Science that gave him a sabbatical semester. Uustalu was supported by the ERDF funded Estonian national CoE project EXCS and ICT national programme project Coinduction (the latter paid also Capretta’s visit to Tallinn), the Estonian Science Foundation grant no. 9475 and the Estonian Ministry of Education and Research institutional research grant no. IUT-3313.

References 1. Abbott, M., Altenkirch, T., Ghani, N.: Containers: constructing strictly positive types. Theoret. Comput. Sci. 342(1), 3–27 (2005) 2. Ad´ amek, J., L¨ ucke, D., Milius, S.: Recursive coalgebras of finitary functors. Theoret. Inform. Appl. 41(4), 447–462 (2007)

106

V. Capretta and T. Uustalu

3. Ad´ amek, J., Milius, S., Moss, L.S., Sousa, L.: Well-pointed coalgebras. Log. Methods Comput. Sci. 9(3), article 2 (2013) 4. Barendregt, H.P., Dekkers, W., Statman, R.: Lambda Calculus with Types: Perspectives in Logic. Cambridge University Press, Cambridge (2013) 5. Berardi, S., Bezem, M., Coquand, T.: On the computational content of the axiom of choice. J. Symb. Log. 63(2), 600–622 (1998) 6. Berger, U., Oliva, P.: Modified bar recursion. Math. Struct. Comput. Sci. 16(2), 163–183 (2006) 7. Capretta, V., Uustalu, T., Vene, V.: Recursive coalgebras from comonads. Inf. Comput. 204(4), 437–468 (2006) 8. Capretta, V., Uustalu, T., Vene, V.: Corecursive algebras: a study of general structured corecursion. In: Oliveira, M.V.M., Woodcock, J. (eds.) SBMF 2009. LNCS, vol. 5902, pp. 84–100. Springer, Heidelberg (2009) 9. Dummett, M.: Elements of Intuitionism, 2nd edn. Oxford Science Publications, Oxford (2000) 10. Escard´ o, M.H., Oliva, P.: Selection functions, bar recursion and backward induction. Math. Struct. Comput. Sci. 20(2), 127–168 (2010) 11. Escard´ o, M.H., Xu, C.: The inconsistency of a Brouwerian continuity principle with the Curry-Howard interpretation. In: Altenkirch, T. (ed.) 13th International Conference on Typed Lambda Calculi and Applications, TLCA 2015, Leibniz International Proceedings in Informatics, vol. 38, pp. 153–164. Dagstuhl Publishing, Saarbr¨ ucken (2015) 12. Gambino, N., Hyland, M.: Wellfounded trees and dependent polynomial functors. In: Berardi, S., Coppo, M., Damiani, F. (eds.) TYPES 2003. LNCS, vol. 3085, pp. 210–225. Springer, Heidelberg (2004) 13. Ghani, N., Hancock, P., Pattinson, D.: Continuous functions on final coalgebras. Electron. Notes Theoret. Comput. Sci. 164(1), 141–155 (2006) 14. Hancock, P., Pattinson, D., Ghani, N.: Representations of stream processors using nested fixed points. Log. Methods Comput. Sci. 5(3), article 9 (2009) 15. Jacobs, B.: The temporal logic of coalgebras via Galois algebras. Math. Struct. Comput. Sci. 12(6), 875–903 (2002) 16. Kleene, S., Vesley, R.: The Foundations of Intuitionistic Mathematics: Especially in Relation to Recursive Functions. North-Holland, Amsterdam (1965) 17. Nakata, K., Uustalu, T., Bezem, M.: A proof pearl with the fan theorem and bar induction. In: Yang, H. (ed.) APLAS 2011. LNCS, vol. 7078, pp. 353–368. Springer, Heidelberg (2011) 18. Osius, G.: Categorical set theory: a characterization of the category of sets. J. Pure Appl. Algebra 4(1), 79–119 (1974) 19. Spector, C.: Provably recursive functionals of analysis: a consistency proof of analysis by an extension of principles in current intuitionistic mathematics. In: Dekker, F.D.E. (ed.) Recursive Function Theory: Proceedings of Symposia in Pure Mathematics. vol. 5, pp. 1–27. American Mathematical Society, Providence, RI (1962) 20. Taylor, P.: Intuitionistic sets and ordinals. J. Symb. Logic 61(3), 705–744 (1996) 21. Taylor, P.: Towards a unified treatment of induction, I: The general recursion theorem. Manuscript (1996) 22. Taylor, P.: Practical Foundations of Mathematics. Cambridge Studies in Advanced Mathematics. Cambridge University Press, Cambridge (1999) 23. Troelstra, A.: Choice Sequences. Clarendon Press, Oxford (1977)

A New Foundation for Finitary Corecursion The Locally Finite Fixpoint and Its Properties Stefan Milius1 , Dirk Pattinson2 , and Thorsten Wißmann1(B) 1

Friedrich-Alexander-Universit¨ at Erlangen-N¨ urnberg, Erlangen, Germany [email protected] 2 The Australian National University, Canberra, Australia

Abstract. This paper contributes to a theory of the behaviour of “finitestate” systems that is generic in the system type. We propose that such systems are modeled as coalgebras with a finitely generated carrier for an endofunctor on a locally finitely presentable category. Their behaviour gives rise to a new fixpoint of the coalgebraic type functor called locally finite fixpoint (LFF). We prove that if the given endofunctor preserves monomorphisms then the LFF always exists and is a subcoalgebra of the final coalgebra (unlike the rational fixpoint previously studied by Ad´ amek, Milius and Velebil). Moreover, we show that the LFF is characterized by two universal properties: 1. as the final locally finitely generated coalgebra, and 2. as the initial fg-iterative algebra. As instances of the LFF we first obtain the known instances of the rational fixpoint, e.g. regular languages, rational streams and formal power-series, regular trees etc. And we obtain a number of new examples, e.g. (realtime deterministic resp. non-deterministic) context-free languages, constructively S-algebraic formal power-series (and any other instance of the generalized powerset construction by Silva, Bonchi, Bonsangue, and Rutten) and the monad of Courcelle’s algebraic trees.

1

Introduction

Coalgebras capture many types of state based system within a uniform and mathematically rich framework [39]. One outstanding feature of the general theory is final semantics which gives a fully abstract account of system behaviour. For example, coalgebraic modelling of deterministic automata (without a finiteness restriction on state sets) yields the set of all formal languages as a final model, and restricting to finite automata one precisely obtains the regular languages [38]. This correspondence has been generalized to locally finitely presentable categories [8,20], where finitely presentable objects play the role of finite sets, leading to the notion of rational fixpoint that provides final semantics to all models with finitely presentable carrier [30]. It is known that the rational fixpoint is fully abstract (identifies all behaviourally equivalent states) as long as finitely presentable objects agree with finitely generated objects in the base category S. Milius and T. Wißmann—Supported by Deutsche Forschungsgemeinschaft (DFG) under project MI 717/5-1. c Springer-Verlag Berlin Heidelberg 2016  B. Jacobs and C. L¨ oding (Eds.): FOSSACS 2016, LNCS 9634, pp. 107–125, 2016. DOI: 10.1007/978-3-662-49630-5 7

108

S. Milius et al.

[12, Proposition 3.12]. While this is the case in some categories (e.g. sets, posets, graphs, vector spaces, commutative monoids), it is currently unknown in other base categories that are used in the construction of system models, for example in idempotent semirings (used in the treatment of context-free grammars [43]), in algebras for the stack monad (used for modelling configurations of stack machines [23]); or it even fails, for example in the category of finitary monads on sets (used in the categorical study of algebraic trees [7]), or Eilenberg-Moore categories for a monad in general (the target category of generalized determinization [41], in which the above examples live). Coalgebras over a category of EilenbergMoore algebras over Set in particular provide a paradigmatic setting: automata that describe languages beyond the regular languages consist of a finite state set, but their transitions produce side effects such as the manipulation of a stack. These can be described by a monad, so that the (infinite) set of system states (machine states plus stack content) is described by a free algebra (for that monad) that is generated by the finite set of machine states. This is formalized by the generalized powerset construction [41] and interacts nicely with the coalgebraic framework we present. Technically, the shortcoming of the rational fixpoint is due to the fact that finitely presentable objects are not closed under quotients, so that the rational fixpoint itself may fail to be a subcoalgebra of the final coalgebra and so identifies too little behaviour. The main conceptual contribution of this paper is the insight that also in cases where finitely presentable and finitely generated do not agree, the locally finite fixpoint provides a fully abstract model of finitely generated behaviour. We give a construction of the locally finite fixpoint, and support our claim both by general results and concrete examples: we show that under mild assumptions, the locally finite fixpoint always exists, and is indeed a subcoalgebra of the final coalgebra. Moreover, we give a characterization of the locally finite fixpoint as the initial iterative algebra. We then instantiate our results to several scenarios studied in the literature. First, we show that the locally finite fixpoint is universal (and fully abstract) for the class of systems produced by the generalized powerset construction over Set: every determinized finite-state system induces a unique homomorphism to the locally finite fixpoint, and the latter contains precisely the finite-state behaviours. Applied to the coalgebraic treatment of context-free languages, we show that the locally finite fixpoint yields precisely the context-free languages, and realtime deterministic context-free languages, respectively, when modelled using algebras for the stack monad of [23]. For context-free languages weighted in a semiring S, or equivalently for constructively S-algebraic power series [36], the locally finite fixpoint comprises precisely those, by phrasing the results of Winter et al. [44] in terms of the generalized powerset construction. Our last example shows the applicability of our results beyond categories of Eilenberg-Moore algebras over Set, and we characterize the monad of Courcelle’s algebraic trees over a signature [7,16] as the locally finite fixpoint of an associated functor (on a category of monads), solving an open problem of [7]. The work presented here is based on the third author’s master thesis in [45]. Most proofs are omitted; they can be found in the full version [33] of our paper.

A New Foundation for Finitary Corecursion

2

109

Preliminaries and Notation

Locally Finitely Presentable Categories. A filtered colimit is the colimit of a diagram D → C where D is filtered (every finite subdiagram has a cocone in D) and directed if D is additionally a poset. Finitary functors preserve filtered (equivalently directed) colimits. Objects C ∈ C are finitely presentable (fp) if the hom-functor C(C, −) preserves filtered (equivalently directed) colimits, and finitely generated (fg) if C(C, −) preserves directed colimits of monos (i.e. colimits of directed diagrams where all connecting morphisms are monic). Clearly any fp object is fg, but not vice versa. Also, fg objects are closed under strong epis (quotients) which fails for fp objects in general. A cocomplete category is locally finitely presentable (lfp) if the full subcategory Cfp of finitely presentable objects is essentially small, i.e. is up to isomorphism only a set, and every object C ∈ C is a filtered colimit of a diagram in Cfp . We refer to [8, 20] for further details. It is well known that the categories of sets, posets and graphs are lfp with finitely presentable objects precisely the finite sets, posets, graphs, respectively. The category of vector spaces is lfp with finite-dimensional spaces being fp. Every finitary variety is lfp (i.e. an equational class of algebras induced by finitearity operations or equivalently the Eilenberg-Moore category for a finitary SetMonad, see Sect. 4.1 later). The finitely generated objects are the finitely generated algebras, and finitely presentable objects are algebras specified by finitely many generators and relations. This includes the categories of groups, monoids, (idempotent) semirings, semi-modules, etc. Every lfp category has mono/strong epi factorization [8, Proposition 1.16], i.e. every f factors as f = m · e with m mono (denoted by ), e strong epi (denoted by ), and we call the domain Im(f ) of e the image of f . Any strong epi e has the diagonal fill-in property, i.e. m · g = h · e with m mono and e strong epi gives a unique d such that m · d = h and g = d · e. Coalgebras. If H : C → C is an endofunctor, H-coalgebras are pairs (C, c) with c : C → HC, and C is the carrier of (C, c). Homomorphisms f : (C, c) → (D, d) are maps f : C → D such that Hf · c = d · f . This gives a category denoted by CoalgH. If its final object exists then this final H-coalgebra (νH, τ ) represents a canonical domain of behaviours of H-typed systems, and induces for each (C, c) a unique homomorphism, denoted by c† , giving semantics to the system (C, c). The final coalgebra always exists provided C is lfp and H is finitary. The forgetful functor CoalgH → C creates colimits and reflects monos and epis. A morphism f in CoalgH is mono-carried (resp. epi-carried ) if the underlying morphism in C is monic (resp. epic). Strong epi/mono factorizations lift from C to CoalgH whenever H preserves monos yielding epi-carried/mono-carried factorizations. A directed union of coalgebras is the colimit of a directed diagram in CoalgH where all connecting morphisms are mono-carried. The Rational Fixpoint. For C lfp and H : C → C finitary let Coalgfp H denote the full subcategory of CoalgH of coalgebras with fp carrier, and Coalglfp H the full subcategory of CoalgH of coalgebras that arise as filtered colimits of coalgebras with fp carrier [30, Corollary III.13]. The coalgebras in Coalglfp H are

110

S. Milius et al.

called lfp coalgebras and for C = Set those are precisely the locally finite coalgebras (i.e. those coalgebras where every element is contained in a finite subcoalgebra). The final lfp coalgebra exists and is the colimit of the inclusion Coalgfp H → CoalgH, and it is a fixpoint of H (see [6]) called the rational fixpoint of H. Here are some examples: the rational fixpoint of a polynomial set functor associated to a finitary signature Σ is the set of rational Σ-trees [6], i.e. finite and infinite Σ-trees having, up to isomorphism, finitely many subtrees only, and one obtains rational weighted languages for Noetherian semirings S for a functor on the category of S-modules [12], and rational λ-trees for a functor on the category of presheaves on finite sets [2] or for a related functor on nominal sets [34]. If the classes of fp and fg objects coincide in C, then the rational fixpoint is a subcoalgebra of the final coalgebra [12, Theorem 3.12]. This is the case in the above examples, but not in general, see [12, Example 3.15] for a concrete example where the rational fixpoint does not identify behaviourally equivalent states. Conversely, even if the classes differ, the rational fixpoint can be a subcoalgebra, e.g. for any constant functor. Iterative Algebras. If H : C → C is an endofunctor, an H-algebra (A, a : HA → A) is iterative if every flat equation morphism e : X → HX + A where X is an fp object has a unique solution, i.e. if there exists a unique e† : X → A such that e† = [a, idA ] · (He† + idA ) · e. The rational fixpoint is also characterized as the initial iterative algebra [6] and is the starting point of the coalgebraic approach to Elgot’s iterative theories [18] and to the iteration theories of Bloom ´ and Esik [3,4, 6,11].

3

The Locally Finite Fixpoint

The locally finite fixpoint can be characterized similarly to the rational fixpoint, but with respect to coalgebras with finitely generated (not finitely presentable) carrier. We show that the locally finite fixpoint always exists, and is a subcoalgebra of the final coalgebra, i.e. identifies all behaviourally equivalent states. As a consequence, the locally finite fixpoint provides a fully abstract notion of finitely generated behaviour. From now on, we rely on the following: Assumption 3.1. Throughout the rest of the paper we assume that C is an lfp category and that H : C → C is finitary and preserves monomorphisms. As for the rational fixpoint, we denote the full subcategory of CoalgH comprising all coalgebras with finitely generated carrier by Coalgfg H and have the following notion of locally finitely generated coalgebra. x

Definition 3.2. A coalgebra X − → HX is called locally finitely generated (lfg) if for all f : S → X with S finitely generated, there exist a coalgebra p : P → HP in Coalgfg H, a coalgebra morphism h : (P, p) → (X, x) and some f  : S → P such that h · f  = f . Coalglfg H ⊆ CoalgH denotes the full subcategory of lfg coalgebras.

A New Foundation for Finitary Corecursion

111

Equivalently, one can characterize lfg coalgebras in terms of subobjects and subcoalgebras, making it a generalization of of local finiteness in Set, i.e. the property of a coalgebra that every element is contained in a finite subcoalgebra. x

f

X , there → HX is an lfg coalgebra iff for all fg subobjects S Lemma 3.3. X −  exist a subcoalgebra h : (P, p)  (X, x) and a mono f : S  P with h · f  = f , i.e. S is a subobject of P . Proof. (⇒) Given some mono f : S  X, factor the induced h into some strong epi-carried and mono-carried homomorphisms and use that fg objects are closed under strong epis. (⇐) Factor f : S → X into an epi and a mono g : Im(f )  X and use the diagonal fill-in property for g.   Evidently all coalgebras with finitely generated carriers are lfg. Moreover, lfg coalgebras are precisely the filtered colimits of coalgebras from Coalgfg H. Proposition 3.4. Every filtered colimit of coalgebras from Coalgfg H is lfg. Proof (Sketch). One first proves that directed unions of coalgebras from Coalgfg H are lfg. Now given a filtered colimit ci : Xi → C where Xi are coalgebras in Coalgfg H, one epi-mono factorizes every colimit injection:

ci = ( Xi ei Ti mi C ). Using the diagonalization of the factorization one sees that the Ti form a directed diagram of subobjects of C. Furthermore C is the   directed union of the Ti and therefore an lfg coalgebra as desired. Proposition 3.5. Every lfg coalgebra (X, x) is a directed colimit of its subcoalgebras from Coalgfg H. Proof. Recall from [8, ProofIofTheorem 1.70] that X is the colimit of the diagram of all its finitely generated subobjects. Now the subdiagram given by all subcoalgebras of X is cofinal. Indeed, this follows directly from the fact that (X, x) is an lfg coalgebra: for every subobject S  X, S fg, we have a subcoalgebra of   (X, x) in Coalgfg H containing S. Corollary 3.6. The lfg coalgebras are precisely the filtered colimits, or equivalently directed unions, of coalgebras with fg carrier. As a consequence, a coalgebra is final in Coalglfg F if there is a unique morphism from every coalgebra with finitely generated carrier. Proposition 3.7. An lfg coalgebra L is final in Coalglfg H iff for every for every coalgebra X in Coalgfg H there exists a unique coalgebra morphism from X to L. The proof is analogous to [30, Theorem 3.14]; the full argument can be found in [33]. Cocompleteness of C ensures that the final lfg coalgebra always exists. Theorem 3.8. The category Coalglfg H has a final object, and the final lfg coalgebra is the colimit of the inclusion Coalgfg H → Coalglfg H.

112

S. Milius et al.

Proof. By Corollary 3.6, the colimit of the inclusion Coalgfg H → Coalglfg H is the same as the colimit of the entire Coalglfg H. And the latter is clearly the final   object of Coalglfg H. This theorem provides a construction of the final lfg coalgebra collecting precisely the behaviours of the coalgebras with fg carriers. In the following we shall show that this construction does indeed identify precisely behaviourally equivalent states, i.e. the final lfg coalgebra is always a subcoalgebra of the final coalgebra. Just like fg objects are closed under quotients – in contrast to fp objects – we have a similar property of lfg coalgebras: Lemma 3.9. Lfg coalgebras are closed under strong quotients, i.e. for every strong epi carried coalgebra homomorphisms X  Y, if X is lfg then so is Y. The failure of this property for lfp coalgebras is the reason why the rational fixpoint is not necessarily a subcoalgebra of the final coalgebra and in particular the rational fixpoint in [12, Example 3.15] is an lfp coalgebra for which the property fails. Theorem 3.10. The final lfg H-coalgebra is a subcoalgebra of the final Hcoalgebra. Proof. Let (L, ) be the final lfg coalgebra. Consider the unique coalgebra morphism L → νH and take its factorization: id

(L, )

e i



(I, i)

m

(νH, τ ) ,

with e strong epi in C.

By Lemma 3.9, I is an lfg coalgebra and so by finality of L we have the coalgebra morphism i† such that idL = i† · e. It follows that e is monic and therefore an iso.   In other words, the final lfg H-coalgebra collects precisely the finitely generated behaviours from the final H-coalgebra. We now show that the final lfg coalgebra is a fixpoint of H which hinges on the following: c

Hc

→ HC, the coalgebra HC −−→ HHC is lfg. Lemma 3.11. For any lfg coalgebra C − Proof. Consider f : S → HC with S finitely generated. As C is lfp we know that HC is the colimit of its fg subobjects, and so f : S → HC factors through some subobject inq : Q  HC with Q fg and f = inq · f  . On the other hand, (C, c) is lfg, i.e. the directed union of its subcoalgebras from Coalgfg H. Then, c since H is finitary and mono-preserving, HC − → HHC is also a directed union Hp and the morphism inq : Q → HC factors through some HP −−→ HHP with (P, p) ∈ Coalgfg H via inp : (P, p)  (C, c), i.e. Hinp · q = inq . Finally, we can construct a coalgebra with fg carrier [q,p]

Hinr

Q + P −−−→ HP −−−→ H(Q + P ) and a coalgebra homomorphism Hinp · [q, p] : Q + P → HC. In the diagram

A New Foundation for Finitary Corecursion

S

f

Hc

HC

HHC

Hinp f



in q q

Q

inl

HHinp Hp

HP

HHP

H(Hinp ·[q,p])

H[q,p]

[q,p]

Q+P

113

[q,p]

HP

Hinr

H(Q + P )

every part trivially commutes, so Hinp · [q, p] is the desired homomorphism.   So with a proof in virtue to Lambek’s Lemma [28, Lemma 2.2], we obtain the desired fixpoint: Theorem 3.12. The carrier of the final lfg H-coalgebra is a fixpoint of H. We denote the above fixpoint by (ϑH, ) and call it the locally finite fixpoint (LFF) of H. In particular, the LFF always exists under Assumption 3.1, providing a finitary corecursion principle. Remark 3.13. As we mentioned in the introduction the rational fixpoint of the finitary functor H is the initial iterative algebra for H. A similar algebraic characterization is possible for the LFF. One simply replaces the fp object X in the definition of a flat equation morphism by an fg object to obtain the notion of an fg-iterative algebra. Theorem 3.14. The LFF is the initial fg-iterative H-algebra. For details, see the full version [33] of our paper or [45]. Relation to the Rational Fixpoint. There are examples, where the rational fixpoint is not a subcoalgebra of the final coalgebra. In categories, where fp and fg objects coincide, the rational fixpoint and the LFF coincide as well (cf. the respective colimit-construction in Sect. 2 and Theorem 3.8). In this section we will see, under slightly stronger assumptions, that fg-carried coalgebras are quotients of fp-carried coalgebras, and in particular the locally finite fixpoint is a quotient of the rational fixpoint: namely its image in the final coalgebra. Assumption 3.15. In addition to Assumption 3.1, assume that in the base category C, every finitely presentable object is a strong quotient of a finitely presentable strong epi projective object and that the endofunctor H also preserves strong epis. The condition that every fg object is the strong quotient of a strong epi projective often is phrased as having enough strong epi projectives [14]. This assumption is apparently very strong but still is met in many situations: Example 3.16 – In categories in which all (strong) epis are split, every object is projective and any endofunctor preserves epis, e.g. in Set or VecK .

114

S. Milius et al.

– In the category of finitary endofunctors Funf (Set), all polynomial functors are projective. The finitely presentable functors are quotients of polynomial functors HΣ , where Σ is a finite signature. – In the Eilenberg-Moore category SetT for a finitary monad T , strong epis are surjective T -algebra homomorphisms, and thus preserved by any endofunctor. In SetT , every free algebra T X is projective; this is easy to see using the projectivity of X in Set. Every finitely generated object of SetT is a strong quotient of some free algebra T X with X finite. For more precise definitions, see Sect. 4.1 later. Proposition 3.17. Every coalgebra in Coalgfg H is a strong quotient of a coalgebra with finitely presentable carrier. Theorem 3.18. ϑH is the image of the rational fixpoint H in the final coalgebra. e

m

Proof. Consider the factorization ( H, r)  (B, b)  (νH, τ ). Since H is the colimit of all fp carried H-coalgebras it is an lfg coalgebra by Proposition 3.4 using that fp objects are also fg. Hence, by Lemma 3.9 the coalgebra B is lfg, too. By Proposition 3.7 it now suffices to show that from every (X, x) ∈ Coalgfg H there exists a unique coalgebra morphism into (B, b). Given (X, x) in Coalgfg H, it is the quotient q : (P, p)  (X, x) of an fp-carried coalgebra by Proposition 3.17. Hence, we obtain a unique coalgebra morphism p† : (P, p) → ( H, r). By finality of νH, we have m · e · p† = x† · q (with x† : (X, x) → (νH, τ )). So the diagonal fill-in property induces a homomorphism (X, x) → (B, b), being the only homomorphism (X, x) → (B, b) by the finality of νH and because m is monic.  

4

Instances of the Locally Finite Fixpoint

We will now present a number of instances of the LFF. First note, that all the known instances of the rational fixpoint (see e.g. [6, 12, 30] are also instances of the locally finite fixpoint, because in all those cases the fp and fg objects coincide. For example, the class of regular languages is the rational fixpoint of 2 × (−)Σ on Set. In this section, we will study further instances of the LFF that are most likely not instances of the rational fixpoint and which – to the best of our knowledge – have not been characterized by a universal property yet: 1. Behaviours of finite-state machines with side-effects as considered by the generalized powerset construction (cf. Sect. 4.1), particularly the following. (a) Deterministic and ordinary context-free languages obtained as the behaviours of deterministic and non-deterministic stack-machines, respectively. (b) Constructively S-algebraic formal power series, i.e. the “context-free” subclass of weighted languages with weights from a semiring S, yielded from weighted context-free grammars. 2. The monad of Courcelle’s algebraic trees.

A New Foundation for Finitary Corecursion

4.1

115

Generalized Powerset Construction

The determinization of a non-deterministic automaton using the powerset construction is an instance of a more general framework, described by Silva, Bonchi, Bonsangue, and Rutten [41] based on an observation by Bartels [10] (see also Jacobs [26]). In that generalized powerset construction, an automaton with sideeffects is turned into an ordinary automaton by internalizing the side-effects in the states. The LFF interacts well with this construction, because it precisely captures the behaviours of finite-state automata with side effects. The notion of side-effect is formalized by a monad, which induces the category, in which the LFF is considered. In the following we assume that readers are familiar with monads and Eilenberg-Moore algebras (see e.g. [29] for an introduction). For a monad T on C we denote by C T the category of Eilenberg-Moore algebras. Recall from [8, Corollary 2.75] that if C is lfp (in most of our examples C is Set) and T is finitary then C T is lfp, too, and for every fp object X the free Eilenberg-Moore algebra T X is fp in C T . In all the examples we consider below, the classes of fp and fg objects either provably differ or it is still unknown whether these classes coincide. Example 4.1. In Sects. 4.4 and 4.5 we are going to make use of Moggi’s exception monad transformer (see e.g. [15]). Let us recall that for a fixed object E, the finitary functor (−) + E together with the unit ηX = inl : X → X + E and multiplication μX = idX + [idE , idE ] : X + E + E → X + E form a finitary monad, the exception monad. Its algebras are E-pointed objects, i.e. objects X, together with a morphism E → X, and homomorphisms are morphisms preserving the pointing. So the induced Eilenberg-Moore category is just the slice category C (−)+E ∼ = E/C. Now, given any monad T we obtain a new monad T (− + E) with obvious unit and multiplication. An Eilenberg-Moore algebra for T (− + E) consists of an Eilenberg-Moore algebra for T and an E-pointing, and homomorphisms are T -algebra homomorphisms preserving the pointing [25]. Now an automaton with side-effects is modelled as an HT -coalgebra, where T is a finitary monad on C providing the type of side-effect. For example, for HX = 2×X Σ , where Σ is an input alphabet, 2 = {0, 1} and T the finite powerset monad on Set, HT -coalgebras are non-deterministic automata. However, the coalgebraic semantics using the final HT -coalgebra does not yield the usual language semantics of non-deterministic automata. To obtain this one considers the final coalgebra of a lifting of H to C T . Denote by U : C T → C the canonical forgetful functor. Definition 4.2. For a functor H : C → C and a monad T : C → C, a lifting of H is a functor H T : C T → C T such that H · U = U · H T . If such a (not necessarily unique) lifting exists, the generalized powerset construction transforms an HT -coalgebra into a H T -coalgebra on C T : For a coalgebra x : X → HT X, HT X carries an Eilenberg-Moore algebra, and one uses

116

S. Milius et al.

freeness of the Eilenberg-Moore algebra T X to obtain a canonical T -algebra homomorphism x : (T X, μT ) → H T (T X, μ†T ). The coalgebraic language semanηX x tics of (X, x) is then given by X −−→ T X −−→νH T , i.e. by composing the unique coalgebra morphism induced by x with ηX . This construction yields a funcx → HT X to x and tor T  : Coalg(HT ) → CoalgH T mapping coalgebras X − homomorphisms f to T f (see e.g. [12, ProofofLemma 3.27] for a proof). Now our aim is to show that the LFF of H T characterizes precisely the coalgebraic language semantics of all fp-carried HT -coalgebras. As the right adjoint U preserves monos and is faithful, we know that H T preserves monos, and as T is finitary, filtered colimits in C T are created by the forgetful functor to C, and we therefore see that H T is finitary. Thus, by Theorem 3.8, ϑH T exists and is a subcoalgebra of νH T . By [37] and [10, Corollary 3.4.19], we know that νH T is carried by νH equipped with a canonical algebra structure. Now let us turn to the desired characterization of ϑH T . Formally, the coalgebraic language semantics of all fp-carried HT -coalgebras is collected by forming T

U

→ CoalgH. the colimit k : K → HK of the diagram Coalgfg HT −→ CoalgH T − This coalgebra K is not yet a subcoalgebra of νH (for C = Set that means, not all behaviourally equivalent states are identified in K), but taking its image in νH we obtain the LFF: Proposition 4.3. The image (I, i) of the unique coalgebra morphism k† : K → νH T is precisely the locally finite fixpoint of the lifting H T . One can also directly take the union of all desired behaviours, for C = Set: Theorem 4.4. The locally finite fixpoint of the lifting H T comprises precisely the images of determinized HT -coalgebras:   T x† [T X] = x† · ηX [X] ⊆ νH T . (1) ϑH T = x:X→HT X X finite

x:X→HT X X finite

This result suggests that the locally finite fixpoint is the right object to consider in order to represent finite behaviour. We now instantiate the general theory with examples from the literature to characterize several well-known notions as LFF. 4.2

The Languages of Non-deterministic Automata

Let us start with a simple standard example. We already mentioned that nondeterministic automata are coalgebras for the functor X → 2 × Pf (X)Σ . Hence they are HT -coalgebras for H = 2 × (−)Σ and T = Pf the finite powerset monad on Set. The above generalized powerset construction then instantiates as the usual powerset construction that assigns to a given non-deterministic automaton its determinization. Now note that the final coalgebra for H is carried by the set L = P(Σ ∗ ) of all formal languages over Σ with the coalgebra structure given by o : L → 2 with

A New Foundation for Finitary Corecursion

117

o(L) = 1 iff L contains the empty word and t : L → LΣ with t(L)(s) = {w | sw ∈ L} the left language derivative. The functor H has a canonical lifting H T on the Eilenberg-Moore category of Pf , viz. the category of join semi-lattices. The final coalgebra νH T is carried by all formal languages with the join semilattice structure given by union and ∅ and with the above coalgebra structure. Furthermore, the coalgebraic language semantics of x : X → HT X assigns to every state of the non-deterministic automaton X the language it accepts. Observe that join semi-lattices form a so-called locally finite variety, i.e. the finitely presentable algebras are precisely the finite ones. Hence, Theorem 4.4 states that the LFF of H T is precisely the subcoalgebra of νH T formed by all languages accepted by finite NFA, i.e. regular languages. Note that in this example the LFF and the rational fixpoint coincide since both fp and fg join semi-lattices are simply the finite ones. Similar characterizations of the coalgebraic language semantics of finite coalgebras follow from Theorem 4.4 in other instances of the generalized powerset construction from [41] (cf. e.g. the treatment of the behaviour of finite weighted automata in [12]). We now turn to examples that, to the best of our knowledge, cannot be treated using the rational fixpoint. 4.3

The Behaviour of Stack Machines

Push-down automata are finite state machines with infinitely many configurations. It is well-known that deterministic and non-deterministic pushdown automata recognize different classes of context-free languages. We will characterize both as instances of the locally finite fixpoint, using the results from [23] on stack machines, which can push or read multiple elements to or from the stack in a single transition, respectively. That is, a transition of a stack machine in a certain state consists of reading an input character, going to a successor state based on the stack’s topmost elements and of modifying the topmost elements of the stack. These stack operations are captured by the stack monad. Definition 4.5 (Stack monad, [22, Proposition 5]). For a finite set of stack ∗ symbols Γ , the stack monad is the submonad T of the store monad (− × Γ ∗ )Γ ∗ Γ∗ ∼ Γ∗ ∗ Γ∗ for which the elements r, t of T X ⊆ (X × Γ ) = X × (Γ ) satisfy the following restriction: there exists k depending on r, t such that for every w ∈ Γ k and u ∈ Γ ∗ , r(wu) = r(w) and t(wu) = t(w)u. Note that the parameter k gives a bound on how may of the topmost stack cells the machine can access in one step. Using the stack monad, stack machines are HT -coalgebras, where H = B × (−)Σ is the Moore automata functor for the finite input alphabet Σ and the set B of all predicates mapping (initial) stack configurations to output values from ∗ 2, taking only the topmost k elements into account: B = {p ∈ 2Γ | ∃k ∈ N0 : ∗ ∀w, u ∈ Γ ∗ , |w| ≥ k : p(wu) = p(w)} ⊆ 2Γ . ∗ The final coalgebra νH is carried by B Σ which is (modulo power laws) a set of predicates, mapping stack configurations to formal languages. Goncharov

118

S. Milius et al.

et al. [23] show that H lifts to SetT and conclude that finite-state HT -coalgebras match the intuition of deterministic pushdown automata without spontaneous transitions. The languages accepted by those automata are precisely the realtime deterministic context-free languages; this notion goes back to Harrison and Havel [24]. We obtain the following, with γ0 playing the role of an initial symbol on the stack: Theorem 4.6. The locally finite fixpoint of H T is carried by the set of all maps ∗ f ∈ B Σ such that for any fixed γ0 ∈ Γ , {w ∈ Σ ∗ | f (w)(γ0 ) = 1} is a real-time deterministic context-free language. Proof. By [23, Theorem 5.5], a language L is a real-time deterministic context-free language iff there exists some x : X → HT X, X finite, with its determinization T (s) ∈ x : T X → HT X and there exist s ∈ X and γ0 ∈ Γ such that f = x† · ηX Σ∗ ∗   B and f (w)(γ0 ) = 1 for all w ∈ Σ . The rest follows by (1). Just as for pushdown automata, the expressiveness of stack machines increases when equipping them with non-determinism. Technically, this is done by considering the non-deterministic stack monad T  , i.e. T  denotes a submonad of the ∗ non-deterministic store monad Pf (− × Γ ∗ )Γ , as described in [23, Sect. 6]. In the non-deterministic setting, a similar property holds, namely that the determinized HT  -coalgebras with finite carrier describe precisely the context-free languages [23, Theorem 6.5]. Combine this with (1): 

Theorem 4.7. The locally finite fixpoint of H T is carried by the set of all maps ∗ f ∈ B Σ such that for any fixed γ0 ∈ Γ , {w ∈ Σ ∗ | f (w)(γ0 ) = 1} is a contextfree language. 4.4

Context-Free Languages and Constructively S-Algebraic Power Series

One generalizes from formal (resp. context-free) languages to weighted formal (resp. context-free) languages by assigning to each word a weight from a fixed semiring. More formally, a weighted language – a.k.a. formal power series – over an input alphabet X is defined as a map X ∗ → S, where S is a semiring. The set of all formal power series is denoted by S X

. Ordinary formal languages are formal power series over the boolean semiring B = {0, 1}, i.e. maps X ∗ → {0, 1}. An important class of formal power series is that of constructively S-algebraic formal power series. We show that this class arises precisely as the LFF of the standard functor for deterministic Moore automata H = S × (−)Σ , but on an Eilenberg-Moore category of a Set monad. As a special case, constructively Balgebraic series are the context-free weighted languages and are precisely the LFF of the automata functor in the category of idempotent semirings. The original definition of constructively S-algebraic formal power series goes back to Fliess [19], see also [17]. Here, we use the equivalent coalgebraic characterization by Winter et al. [44]. Let S X ⊆ S X

the subset of those maps, that are 0 for all but finitely many w ∈ X ∗ . If S is commutative, then S − yields a finitary monad and

A New Foundation for Finitary Corecursion

119

thus also T = S − + Σ by Example 4.1. The algebras for S − are associative S-algebras (over the commutative semiring S), i.e. S-modules together with a monoid structure that is a module morphism in both arguments. The algebras for T are Σ-pointed S-algebras. The following notions are special instances of S-algebras. Example 4.8. For S = B = {0, 1}, one obtains idempotent semirings as Balgebras, for S = N semirings, and for S = Z ordinary rings. Winter et al. [44, Proposition 4] show that the final H-coalgebra is carried by S Σ

and that constructively S-algebraic series are precisely those elements of S Σ

that arise as the behaviours of those coalgebra c : X → HS X with finite X, after determinizing them to some c : S X → HS X (see [44, Theorem 23]). However, this determinization is not directly an instance of the generalized powerset construction. We shall show that the same behaviours can be obtained by using the standard generalized powerset construction with an appropriate lifting of H to T -algebras. Having an S-algebra structure on A and a Σ-pointing j : Σ → A we need to define another S-algebra structure and Σ-pointing on HA = S × AΣ . While the S-module structure is just point-wise, we need to take care when multiplying two elements from HA. To this end we first we define the operation [−, −] : S × AΣ → A by   j(b) · δ(b) , [o, δ] := i(o) + b∈Σ

where i : S → A is the canonical map with i(s) = s · 1 with 1 the neutral element of the monoid on A. The idea is that [o, δ] acts like a state with output o and derivation δ. The multiplication on HA = S × AΣ is then defined by   (2) (o1 , δ1 ) ∗ (o2 , δ2 ) := o1 · o2 , a → δ1 (a) · [o2 , δ2 ] + i(o1 ) · δ2 (a) . The Σ-pointing is the obvious: a → (0, a ) where a (a) = 1 and a (b) = 0 for a = b. Lemma 4.9. For any w ∈ A in SetT and any H T -coalgebra structure c : A → H T A, w and [c(w)] are behaviourally equivalent in Set. Given a coalgebra c : X → HS X , Winter et al. [44, Proposition 14] determinize ˆ : S X → HS X with the property that for any v, w ∈ c to some cˆ = ˆ o, δ S X , oˆ(v ∗ w) = oˆ(v) · oˆ(w)

ˆ ∗ w, a) = δ(v, ˆ a) ∗ w + oˆ(v) ∗ δ(w, ˆ and δ(v a),

(3)

and such that cˆ is a S-module homomorphism. However, the generalized powerset construction w.r.t. T yields a coalgebra c : S X + Σ → HS X + Σ . The above property, together with Lemma 4.9 and (2) implies that cˆ and c are essentially the same coalgebra structures:

120

S. Milius et al.

Lemma 4.10. In Set, u ∈ (S X , cˆ) and S inl (u) ∈ (S X + Σ , c ) are behaviourally equivalent. It follows that cˆ† = c† · S inl and thus their images in νH are identical. Hence, a formal power series is constructively S-algebraic iff it is in the image of some c† · S inl , and by (1), iff it is in the locally finite fixpoint of H T . 4.5

Courcelle’s Algebraic Trees

For a fixed signature Σ of so called givens, a recursive pro+ gram scheme (or rps, for short) contains mutually recur+ z sive definitions of new operations ϕ1 , . . . , ϕk (with respec× + tive arities n1 , . . . , nk ). The recursive definition of ϕi may  z × involve symbols from Σ, operations ϕ1 , . . . , ϕk and ni vari.. ables x1 , . . . , xni . The (uninterpreted) solution of an rps  × . is obtained by unravelling these recursive definitions, pro z ducing a possibly infinite Σ-tree over x1 , . . . , xni for each operation ϕi . Figure 1 shows an rps over the signature Fig. 1. Solution of Σ = {/0, ×/2, +/2} and its solution. In general, an alge- ϕ(z) = z + ϕ( × z) braic Σ-tree is a Σ-tree which is definable by an rps over Σ (see Courcelle [16]). Generalizing from a signature to a finitary endofunctor H : C → C on an lfp category, Ad´ amek et al. [7] describe an rps as a coalgebra for a functor Hf on H/Mndf (C), in which objects are finitary H-pointed monads on C, i.e. finitary monads M together with a natural transformation H → M . They introduce the context-free monad C H of H, which is an H-pointed monad that is a subcoalgebra of the final coalgebra for Hf and which is the monad of Courcelle’s algebraic Σ-trees in the special case where C = Set and H is a polynomial functor associated to a signature Σ. We will prove that this monad is the LFF of Hf , and thereby we characterize it by a universal property – solving the open problem in [7]. The setting is again an instance of the generalized powerset construction, but this time with Funf (C) as the base category in lieu of Set. Let C be an lfp category in which the coproduct injections are monic and consider a finitary, mono-preserving endofunctor H : C → C. Denote by Funf (C) the category of finitary endofunctors on C. Then H induces an endofunctor H · (−) + Id on Funf (C), denoted H˙ and mapping an endofunctor V to the functor X → HV X + X. This functor H˙ gets precomposed with a monad on Funf (C) as we now explain. Proposition 4.11 (Free monad, [5, 9]). For a finitary endofunctor H, free H-algebras ϕX : HF H X → F H X exist for all X ∈ C. F H itself is a finitary monad on C, more specifically it is the free monad on H. For example, if H is a polynomial functor associated to a signature Σ, then F H X is the usual term algebra that contains all finite Σ-trees over the set of generators X. Proposition 4.11 implies that H → F H is the object assignment of a monad on Funf (C). The Eilenberg-Moore category of F (−) is easily seen to be Mndf (C), the category of finitary monads on C. Here, fp and fg objects differ, see [45, Sect. 5.4.1] for a proof.

A New Foundation for Finitary Corecursion

121

Similarly as in the case of context-free languages, we will work with the monad E (−) = F H+(−) , so we get H-pointed finitary monads as the E (−) algebras. This category is equivalent to a slice category: the universal property induced by F (−) states, that for any finitary monad B the natural transformations H → B are in one-to-one correspondence with monad morphisms F H → B; so the category H/Mndf (C) of finitary H-pointed monads on C is isomorphic to the slice category F H /Mndf (C). This finishes the description of the base category and we now lift the functor H˙ to this category. Consider an H-pointed monad (B, β : H → C) ∈ H/Mndf (C). By [21], the endofunctor H · B + Id carries a canonical monad structure. Furthermore, we have an obvious pointing inl · Hη B : H → H · B + Id. By [32], this defines an endofunctor on H-pointed monads, Hf : H/Mndf (C) → H/Mndf (C), which is a ˙ In order to verify that Hf is finitary, we first need to know how lifting of H. filtered colimits look in H/Mndf (C). Lemma 4.12. The forgetful U : Mndf (C) → Funf (C) creates filtered colimits. Clearly, the canonical projection functor H/Mndf (C) → Mndf (C) creates filtered colimits, too. Therefore, filtered colimits in the slice category H/Mndf (C) are formed on the level of Funf (C), i.e. object-wise. The functor H˙ is finitary on Funf (C) and thus also its lifting Hf is finitary. So all requirements from Assumption 3.1 are met: we have a finitary endofunctor Hf on the lfp category H/Mndf (C), and by [7, Corollary 2.20] Hf preserves monos since H does. By Theorem 3.8, Hf has a locally finite fixpoint. Remark 4.13. The final Hf -coalgebra is not of much interest, but that of a related functor. Hf generalizes to a functor H : H/Mndc (C) → H/Mndc (C) on H-pointed countably accessible1 monads. For any object X ∈ C, the finitary endofunctor H(−) + X has a final coalgebra; call the carrier T X. Then T is a monad [1], is countably accessible [7] and is the final H-coalgebra [32]. Ad´ amek et al. [7] characterize a (guarded) recursive program scheme as a natural transformation V → H · E V + Id with V fp (in Funf (C)), or equivalently, via the generalized powerset construction w.r.t. the monad E (−) as an Hf -coalgebra on the carrier E V (in Mndf (C)). These Hf -coalgebras on carriers E V where V ∈ Funf (C) is fp form the full subcategory EQ ⊆ CoalgHf . They show two equivalent ways of constructing the monad of Courcelle’s algebraic trees for the case C = Set: as the image of EQcolim in the final coalgebra T of Remark 4.13, and as the colimit of EQ2 , where EQ2 is the closure of EQ under strong quotients. We now provide a third characterization, and show that the monad of Courcelle’s algebraic trees is the locally finite fixpoint of Hf . To this end it suffices to show that EQ2 is precisely the diagram of Hf coalgebras with an fg carrier. This is established with the help of the following two technical lemmas. We now assume that C = Set. 1

A colimit is countably filtered if its diagram has for every countable subcategory a cocone. A functor is countably accessible if it preserves countably filtered colimits.

122

S. Milius et al.

Lemma 4.14. Hf maps strong epis to morphisms carried by strong epi natural transformations. We have the following variation of Proposition 3.17: Lemma 4.15. Any Hf -coalgebra b : (B, β) → Hf (B, β), with B fg, is the strong quotient of a coalgebra from EQ. The proof of Lemma 4.15 makes use of Lemma 4.14 as well as the following properties: – The fp objects in Funf (Set) are the quotients of polynomial functors. – The polynomial functors are projective. That means that for a polynomial functor P and any natural transformation n : K → L with surjective components we have the following property: for every f : P → L there exists f  : P → K with n · f  = f . – Any fg object in H/Mndf (Set) is the quotient of some E V with V fp in Funf (Set) and thus also of some E P with P a polynomial functor. Note that the last property holds because H/Mndf (Set) is an Eilenberg-Moore category and E V is the free Eilenberg-Moore algebra on the fp object V . It follows from Lemma 4.15 that Coalgfg Hf is the same category as EQ2 ; thus their colimits in CoalgHf are isomorphic and we conclude: Theorem 4.16. The locally finite fixpoint of Hf : HΣ /Mndf (Set) → HΣ /Mndf (Set) is the monad of Courcelle’s algebraic trees, sending a set to the algebraic Σ-trees over it.

5

Conclusions and Future Work

We have introduced the locally finite fixpoint of a finitary mono-preserving endofunctor on an lfp category. We proved that this fixpoint is characterized by two universal properties: it is the final lfg coalgebra and the initial fg-iterative algebra for the given endofunctor. And we have seen many instances where the LFF is the domain of behaviour of finite-state and finite-equation systems. In particular all previously known instances of the rational fixpoint are also instances of the LFF, and we have obtained a number of interesting further instances not captured by the rational fixpoint. On a more technical level, the LFF solves a problem that sometimes makes the rational fixpoint hard to apply. The latter identifies behaviourally equivalent states (i.e. is a subcoalgebra of the final coalgebra) if the classes of fp and fg objects coincide. This condition, however, may be false or unknown (and sometimes non-trivial to establish) in a given lfp category. But the LFF always identifies behaviourally equivalent states. There are a number of interesting topics for future work concerning the LFF. First, it should be interesting to obtain further instances of the LFF, e.g. analyzing the behaviour of tape machines [23] may perhaps lead to a description of the

A New Foundation for Finitary Corecursion

123

recursively enumerable languages by the LFF. Second, syntactic descriptions of the LFF are of interest. In works such as [12, 35, 40, 42] Kleene type theorems and axiomatizations of the behaviour of finite systems are studied. Completeness of an axiomatization is then established by proving that expressions modulo axioms form the rational fixpoint. It is an interesting question whether the theory of the LFF we presented here may be of help as a tool for syntactic descriptions and axiomatizations of further system types. As we have mentioned already the rational fixpoint is the starting point for the coalgebraic study of iterative and iteration theories. A similar path could be followed based on the LFF and this should lead to new coalgebraic iteration/recursion principles, in particular in instances such as context-free languages or constructively S-algebraic formal power series. Another approach to more powerful recursive definition principles are abstract operational rules (see [27] for an overview). It has been shown that certain rule formats define operations on the rational fixpoint [13, 31], and it should be investigated whether a similar theory can be developed based on the LFF. Finally, in the special setting of Eilenberg-Moore categories one could base the study of finite systems on free finitely generated algebras (rather than all fp or all fg algebras). Does this give a third fixpoint capturing behaviour of finite state systems with side effects besides the rational fixpoint and the LFF? And what is then the relation between the three fixpoints? Also the parallelism in the technical development between rational fixpoint and LFF indicates that there should be a general theory that is parametric in a class of “finite objects” and that allows to obtain results about the rational fixpoint, the LFF and other possible “finite behaviour domains” as instances.

References 1. Aczel, P., Ad´ amek, J., Milius, S., Velebil, J.: Infinite trees and completely iterative theories: a coalgebraic view. Theoret. Comput. Sci. 300, 1–45 (2003) 2. Ad´ amek, J., Milius, S., Velebil, J.: Semantics of higher-order recursion schemes. Log. Methods Comput. Sci. 7(1:15), 43 (2011) 3. Ad´ amek, J., Milius, S., Velebil, J.: Equational properties of iterative monads. Inf. Comput. 208, 1306–1348 (2010) 4. Ad´ amek, J., Milius, S., Velebil, J.: Elgot theories: a new perspective of the equational properties of iteration. Math. Struct. Comput. Sci. 21(2), 417–480 (2011) 5. Ad´ amek, J.: Free algebras and automata realizations in the language of categories. Comment. Math. Univ. Carolin. 015(4), 589–602 (1974) 6. Ad´ amek, J., Milius, S., Velebil, J.: Iterative algebras at work. Math. Struct. Comput. Sci. 16(6), 1085–1131 (2006) 7. Ad´ amek, J., Milius, S., Velebil, J.: On second-order iterative monads. Theoret. Comput. Sci. 412(38), 4969–4988 (2011) 8. Ad´ amek, J., Rosick´ y, J.: Locally Presentable and Accessible Categories. Cambridge University Press, New York (1994) 9. Barr, M.: Coequalizers and free triples. Math. Z. 116, 307–322 (1970)

124

S. Milius et al.

10. Bartels, F.: On generalized coinduction and probabilistic specification formats: Distributive laws in coalgebraic modelling. Ph.D. thesis, Vrije Universiteit Amsterdam.(2004) ´ 11. Bloom, S.L., Esik, Z.: Iteration Theories: The Equational Logic of Iterative Processes. EATCS Monographs on Theoretical Computer Science. Springer, Heidelberg (1993) 12. Bonsangue, M., Milius, S., Silva, A.: Sound and complete axiomatizations of coalgebraic language equivalence. ACM Trans. Comput. Log. 14((1: 7)), 52.(2013) 13. Bonsangue, M.M., Milius, S., Rot, J.: On the specification of operations on the rational behaviour of systems. In: Luttik, B., Reniers, M.A. (eds.) Proceedings of Combined Workshop on Expressiveness in Concurrency and Structural Operational Semantics (EXPRESS/SOS’12), Electronic Proceedings of Theoretical Computer Science, vol. 89, pp. 3–18.(2012) 14. Borceux, F.: Handbook of Categorical Algebra: Volume 1, Basic Category Theory. Encyclopedia of Mathematics and its Applications. Cambridge University Press, Cambridge (1994) 15. Cenciarelli, P., Moggi, E.: A syntactic approach to modularity in denotational semantic. In: Proceedings of 5th CTCS. CWI Technical report.(1993) 16. Courcelle, B.: Fundamental properties of infinite trees. Theoret. Comput. Sci. 25, 95–169 (1983) 17. Droste, M., Kuich, W., Vogler, H.: Handbook of Weighted Automata, 1st edn. Springer Publishing Company, Berlin (2009) 18. Elgot, C.: Monadic computation and iterative algebraic theories. In: Rose, H.E., Sheperdson, J.C. (eds.) Logic Colloquium 1973, vol. 80, pp. 175–230. NorthHolland Publishers, Amsterdam (1975) 19. Fliess, M.: Sur divers produits de s´eries formelles. Bulletin de la Soci´et´e Math´ematique de France 102, 181–191 (1974) 20. Gabriel, P., Ulmer, F.: Lokal pr¨ asentierbare Kategorien. Lecture Notes in Mathematics, vol. 221. Springer, Heidelberg (1971) 21. Ghani, N., L¨ uth, C., Marchi, F.D.: Monads of coalgebras: rational terms and term graphs. Math. Struct. Comput. Sci. 15, 433–451 (2005) 22. Goncharov, Sergey: Trace Semantics via Generic Observations. In: Heckel, Reiko, Milius, Stefan (eds.) CALCO 2013. LNCS, vol. 8089, pp. 158–174. Springer, Heidelberg (2013) 23. Goncharov, Sergey, Milius, Stefan, Silva, Alexandra: Towards a Coalgebraic Chomsky Hierarchy. In: Diaz, Josep, Lanese, Ivan, Sangiorgi, Davide (eds.) TCS 2014. LNCS, vol. 8705, pp. 265–280. Springer, Heidelberg (2014) 24. Harrison, M.A., Havel, I.M.: Real-time strict deterministic languages. SIAM J. Comput. 1(4), 333–349 (1972) 25. Hyland, M., Plotkin, G., Power, J.: Combining effects: sum and tensor. Theoret. Comput. Sci. 357(1–3), 70–99 (2006) 26. Jacobs, Bart: A Bialgebraic Review of Deterministic Automata, Regular Expressions and Languages. In: Futatsugi, Kokichi, Jouannaud, Jean-Pierre, Meseguer, Jos´e (eds.) Algebra, Meaning, and Computation. LNCS, vol. 4060, pp. 375–404. Springer, Heidelberg (2006) 27. Klin, B.: Bialgebras for structural operational semantics: an introduction. Theoret. Comput. Sci. 412(38), 5043–5069 (2011) 28. Lambek, J.: A fixpoint theorem for complete categories. Math. Z. 103, 151–161 (1968) 29. MacLane, S.: Categories for the Working Mathematician. Graduate Texts in Mathematics, 2nd edn. Springer, New York.(1998)

A New Foundation for Finitary Corecursion

125

30. Milius, S.: A sound and complete calculus for finite stream circuits. In: Proceedings of 25th Annual Symposium on Logic in Computer Science (LICS 2010), pp. 449– 458.(2010) 31. Milius, S., Bonsangue, M.M., Myers, R.S., Rot, J.: Rational operation models. Electron. Notes Theoret. Comput. Sci. 298, 257–282. In: Mislove, M. (ed.) Proceedings of 29th conference on Mathematical Foundations of Programming Science (MFPS XXIX).(2013) 32. Milius, S., Moss, L.S.: The category theoretic solution of recursive program schemes. Theoret. Comput. Sci. 366, 3–59 (2006) 33. Milius, S., Pattinson, D., Wißmann, T.: A new foundation for finitary corecursion: the locally finite fixpoint and its properties (2015).. http://arxiv.org/abs/1601. 01532 34. Milius, S., Wißmann, T.: Finitary corecursion for the infinitary lambda calculus. In: Proceedings of 6th Conference on Algebra and Coalgebra in Computer Science, CALCO 2015. Leibniz International Proceedings in Informatics.(2015) 35. Myers, R.: Rational coalgebraic machines in varieties: Languages, completeness and automatic proofs. Ph.D. thesis, Imperial College London, Department of Computing.(2011) 36. Petre, I., Salomaa, A.: Algebraic systems and pushdown automata. In: Droste, M., Kuich, W., Vogler, H. (eds.) Handbook of Weighted Automata. Monographs in Theoretical Computer Science. An EATCS Series, pp. 257–289.. Springer, Heidelberg (2009) 37. Plotkin, G., Turi, D.: Towards a mathematical operational semantics. In: Proceedings of 12th LICS Conference, pp. 280–291. IEEE Computer Society Press.(1997) 38. Rutten, J.J.M.M.: Automata and Coinduction. In: Sangiorgi, Davide, de Simone, Robert (eds.) CONCUR 1998. LNCS, vol. 1466, pp. 194–218. Springer, Heidelberg (1998) 39. Rutten, J.: Universal coalgebra: a theory of systems. Theoret. Comput. Sci. 249(1), 3–80 (2000) 40. Silva, A., Bonchi, F., Bonsangue, M.M., Rutten, J.J.M.M.: Quantitative kleene coalgebras. Inf. Comput. 209(5), 822–849 (2011) 41. Silva, A., Bonchi, F., Bonsangue, M.M., Rutten, J.: Generalizing determinization from automata to coalgebras. Log. Meth. Comput. Sci. 9(1), 1–27 (2013) 42. Silva, A., Bonsangue, M.M., Rutten, J.: Non-deterministic Kleene coalgebras. Log. Meth. Comput. Sci. 6(3: 23), 39.(2010) 43. Winter, J., Bonsangue, M., Rutten, J.: Coalgebraic characterizations of contextfree languages. Log. Meth. Comput. Sci. 9(3), 1–39 (2013) 44. Winter, J., Bonsangue, M.M., Rutten, J.J.: Context-free coalgebras. J. Comput. Syst. Sci. 81(5), 911–939 (2015) 45. Wißmann, T.: The locally finite fixpoint and its properties. Master’s thesis, Friedrich-Alexander Universit¨ at Erlangen-N¨ urnberg (April 2015). http:// thorsten-wissmann.de/theses/ma-wissmann.pdf

Fixed-Point Elimination in the Intuitionistic Propositional Calculus Silvio Ghilardi1 , Maria Jo˜ ao Gouveia2 , and Luigi Santocanale3(B) 1

Dipartimento di Matematica, Universit` a degli Studi di Milano, Milan, Italy ˆ CEMAT-CIENCIAS, Universidade de Lisboa, 1749-016 Lisboa, Portugal 3 LIF, CNRS UMR 7279, Aix-Marseille Universit´e, Marseille, France [email protected]

2

Abstract. It is a consequence of existing literature that least and greatest fixed-points of monotone polynomials on Heyting algebras—that is, the algebraic models of the Intuitionistic Propositional Calculus—always exist, even when these algebras are not complete as lattices. The reason is that these extremal fixed-points are definable by formulas of the IPC. Consequently, the µ-calculus based on intuitionistic logic is trivial, every µ-formula being equivalent to a fixed-point free formula. We give in this paper an axiomatization of least and greatest fixed-points of formulas, and an algorithm to compute a fixed-point free formula equivalent to a given µ-formula. The axiomatization of the greatest fixed-point is simple. The axiomatization of the least fixed-point is more complex, in particular every monotone formula converges to its least fixed-point by Kleene’s iteration in a finite number of steps, but there is no uniform upper bound on the number of iterations. We extract, out of the algorithm, upper bounds for such n, depending on the size of the formula. For some formulas, we show that these upper bounds are polynomial and optimal.

1

Introduction

In [23] the author proved that, for each formula φ(x) of the Intuitionistic Propositional Calculus, there exists a number n ≥ 0 such that φn (x)—the formula obtained from φ by iterating n times substitution of φ for the variable x—and φn+2 (x) are equivalent in Intuitionistic Logic. This result has, as an immediate corollary, that a syntactically monotone formula φ(x) converges both to its least fixed-point and to its greatest fixed-point in at most n steps. Using a modern notation based on μ-calculi [3], we have μx .φ(x) = φn (⊥) and νx .φ(x) = φn (). These identities also show that a μ-calculus based on Intuitionistic Logic is trivial, every μ-formula being equivalent to a fixed-point free formula. Ruitenberg’s work [23] leaves open how to extract or estimate the least number ρ(φ) such that φρ(φ) (x) = φρ(φ)+2 (x). Yet, our motivations stem from the theory of extremal fixed-points and μ-calculi [3]. In principle, being able to compute or bound Ruitenberg’s number ρ(φ) might end up in an over-approximation of the closure ordinal of φ—the least k such that μx .φ(x) = φk (⊥). For the analogous problem with the greatest fixed-point, we shall see that the least number k such that νx .φ(x) = φk () is bounded by 1, while ρ(φ) might be arbitrarily large. c Springer-Verlag Berlin Heidelberg 2016  B. Jacobs and C. L¨ oding (Eds.): FOSSACS 2016, LNCS 9634, pp. 126–141, 2016. DOI: 10.1007/978-3-662-49630-5 8

Fixed-Point Elimination in the Intuitionistic Propositional Calculus

127

Later in [20], the author gave an independent proof that least fixed-points of monotone formulas are definable within Intuitionistic Logic. His proof relies on semantics methods and on the coding of Intuitionistic Logic into Grzegorczyk’s Logic; the proof was further refined in [21] to encompass the standard coding of Intuitionistic Logic into its modal companion, the logic S4. Curiously, no mention of greatest fixed-points appears in these works. Another relevant source for this paper stem from the discovery that IPC has uniform interpolants [22], often named bisimulation quantifiers. Together with the deduction property of IPC, they give the category of (finitely generated) Heyting algebras—that is, the algebraic models of the Intuitionistic Propositional Calculus—a rather strong structure, axiomatized and studied in [14, 15]. It is possible to argue that in every category with similar properties the extremal fixed-points of monotone formulas are definable. This is possible by using quantified formulas analogous to the one used in [9, Sect. 3] to argue that PDL lacks the uniform interpolation property. In this paper we exploit this idea and the existential bisimulation quantifiers to characterize greatest fixed-points in the Intuitionistic Propositional Calculus. A μ-calculus is a prototypical kind of computational logic, obtained from a base logic or algebraic system by addition of distinct forms of iteration so to increase expressivity. This paper is part of a line of research whose goal is to understand, under a unified perspective, why alternation-depth hierarchies of μ-calculi are degenerate or trivial. A μ-calculus adds to an underlying logicalalgebraic system formal lgfps of formula-terms whose semantic monotonicity can be witnessed at the syntactic level. When addition of extremal fixed-points is iterated, formula-terms with nested extremal fixed-points are generated. The alternation-depth hierarchy [3, Sect. 2.6] of a μ-calculus measures the complexity of a formula-term as a function of the nesting of the different types of fixedpoints, with respect to a fixed class of models. It is well known that fixed-points that are unguarded can be eliminated in the propositional modal μ-calculus [18]. We can rephrase this fact by saying that the alternation-depth hierarchy of the μ-calculus over distributive lattices is trivial, every μ-term being equivalent to a fixed-point free term. A goal of [12] was to understand closely this result and to generalize it. We were able to exhibit equational classes of lattices Dn — with D0 the class of distributive lattices—where the extremal fixed-points can be uniformly computed by iterating a formula-term n+1 times from the bottom/top of the lattice; moreover, we showed that these uniform upper bounds are optimal. The reasons for the degeneracy of the hierarchy can be ultimately found in the structural theory of lattices. As we show in this paper, the situation is quite different when the base for the μ-calculus is Intuitionistic Logic, with its standard models the Heyting algebras. Several ingredients contribute to the existence of a closure ordinal of each formula and to its finiteness. Among them, strongness of the monotone polynomials on Heyting algebras. This means that a monotone polynomial f : H −−→ H over a Heyting algebra H can be considered as a functor enriched over H, when H is consider as a closed category [16]. For some polynomials, existence and finiteness

128

S. Ghilardi et al.

of the closure ordinal is a consequence of being inflating (or expanding) and, on the syntactic level, to a restriction to the use of conjunction that determines a notion of disjunctive formula. As far as the greatest fixed-point is concerned, monotone formulas uniformly converge to it after one step. A key ingredient of the algorithm we present is creation of least fixed-points via the Rolling equation (cf. Lemma 1), a fact already used in [10]. For Intuitionistic Logic and Heyting algebras, where formula-terms can be semantically antitone (i.e. contravariant), existing greatest fixed-points create least fixed-points. The most striking difference with the case of distributive lattices (and with the case of the varieties Dn ) is the absence of a finite uniform upper bound on the closure ordinals, the rate of convergence to the least fixed-point crucially depending on the shape of the formula. As emphasized in [19] for the propositional modal μ-calculus, once a formula is known to be equivalent to some other formula of smaller complexity, we should also be able to effectively compute this second formula. Thus, the fact that the alternation hierarchy is trivial for μ-calculi based on the IPC should not be the end of the story. The main contribution of this paper is to achieve an effective transformation of an intuitionisitc μ-formula into an equivalent fixed-point free intuitionisitc formula. The size of the formula might exhibit an exponential grow during this transformation. Yet, this is mainly due, as usual, to the need of precompiling a formula into an equivalent one in some kind of conjunctive normal form. We might use sharing in substitutions—or introduce the appropriate formalism for approximants to least fixed-points—so that, if we are given an already precompiled formula, then its least fixed-point w.r.t. the variable x is only polynomially bigger than the original formula. For these formulas, we instantiate this claim by explicitly giving a way of computing f (φ) such that μx .φ(x) = φf (φ) (⊥), so that f (φ) is an upper bound to the closure ordinal of φ. In some cases we are able to show that f (φ) is optimal, by exhibiting some formula φ(x) such that φf (φ)−1 (⊥) < μx .φ(x). The paper is structured as follows. We recall in Sect. 2 some elementary facts from fixed-point theory. In Sect. 3 we recall the Intuitionistic Propositional Calculus and introduce the Intuitionistic Propositional μ-Calculus. In Sect. 4 we argue that monotone polynomials are strong and exhibit the interactions between least fixed-points and strong functions. In Sect. 5 we use the existential bisimulation quantifier to argue that monotone polynomials converge to their greatest fixed-point in one step. Section 6 is the core of our paper, where we show ho to eliminate a least fixed-point from a formula. Together with the result in the previous Section, this leads to a procedure to eliminate off the fixedpoints from a IPCμ formula. Finally, in Sect. 7, we show how upper bounds to closure ordinals can be extracted from the procedure elimination of the least fixed-points. In Sect. 8 we present our final remarks.

2

Notation and Elementary Concepts

Let P and Q be posets. A function f : P −−→ Q is monotone if x ≤ y implies f (x) ≤ f (y), for each x, y ∈ P . If f : P −−→ P is a monotone endofunction,

Fixed-Point Elimination in the Intuitionistic Propositional Calculus

129

then x ∈ P is a prefixed-point of f if f (x) ≤ x; we denote by Pref the set of prefixed points of f . Whenever Pref has a least element, we denote it by μ.f . Therefore, μ.f denotes the least prefixed-point of f , whenever it exists. If μ.f exists, then it is a fixed-point of f , necessarily the least one. The notions of least prefixed-point and of least fixed-point coincide on complete lattices or when the least fixed-point is computed by iterating from the bottom of a lattice; for our purposes they are interchangeable, so we shall abuse of language and refer to μ.f as the least fixed-point of f . Dually (and abusing again of language), the greatest fixed-point of f shall be denoted by ν.f . Let us mention few elementary facts from fixed-point theory. Lemma 1. Let P, Q be posets, f : P −−→ Q and g : Q −−→ P be monotone functions. If μ.(g◦f ) exists, then μ.(f ◦g) exists as well and is equal to f (μ.(g◦f )). As we do not work in complete lattices (so we are not ensured that least fixedpoints exist) we express the above statement via the equality μ.(f ◦ g) := f (μ.(g ◦ f )) ,

(Roll)

where the colon emphasizes existence: if the least fixed-point in the expression on the right exists, then this expression is the least fixed-point of f ◦g. Analogous notations will be used later. We endow the product of two posets P and Q with the coordinatewise ordering. Therefore a function f : P × Q −−→ R is monotone if, as a function of two variables, it is monotone in each variable. To deal with least fixed-points of functions of many variables, we use the standard notation: for example, if f : P × P −−→ P is the monotone function f (x, y), then, for a fixed p ∈ P , μx .f (x, p) denotes the least fixed-point of f (x, p). Let us recall that the correspondence p → μx .f (x, p)—noted μx .f (x, y)—is again monotone. Lemma 2. If P is a poset and f : P × P −−→ P is a monotone mapping, then μx .f (x, x) := μx .μy .f (x, y) .

(Diag)

Lemma 3. If P and Q are posets and f, g : P × Q −−→ P × Q is a monotone function, then μ. f, g := μ1 , μ2 , where μ1 = μx .f (x, μy .g(x, y))

3

and

μ2 = μy .g(μ1 , y) .

(Bekic)

The Intuitionistic Propositional µ-Calculus

Formulas of the Intuitionistic Propositional Calculus are generated according to the following grammar: φ ⇒ x |  | φ ∧ φ | ⊥ | φ ∨ φ | φ → φ,

(1)

where x ranges over a countable set X of propositional variables. For the IPC, the formulation of the consequence relation LJ (relating a set of formulas to a formula) goes back to Gentzen’s work on the system LJ [13]. It is well known that Intuitionistic Logic is sound and complete w.r.t. the class of its algebraic models, the Heyting algebras.

130

S. Ghilardi et al.

Definition 1. A Heyting algebra H is a bounded lattice (with least element ⊥ and greatest element ) equipped with a binary operation → such that the following equations hold in H: x ∧ (x → y) = x ∧ y , x → x = ,

x ∧ (y → x) = x , x → (y ∧ z) = (x → y) ∧ (x → z) .

(2)

We can define on any Heyting algebra a partial order by saying that x ≤ y holds when x ∨ y = y. We identify formulas of the IPC with terms of the theory of Heyting algebras, constructed therefore from variables and using the signature , ∧, ⊥, ∨, → . For φ such a formula-term, H a Heyting algebra, and v : X −−→ H a valuation of the propositional variables in H, let us write φv for the result of evaluating the formula in H, starting from the variables. The soundness and completeness theorem of the IPC over Heyting algebras—see e.g. [6]—can then be stated as follows: if Γ is a finite setof formula-terms and φ is a formula-term, then Γ LJ φ holds if and only if γ∈Γ γv ≤ φv holds, in every Heyting algebra H and for every valuation of the propositional variables v : X −−→ H. Given this theorem, we shall often abuse of notation and write ≤ in place of LJ , and the equality symbol = to denote logical equivalence of formulas. We aim at studying extremal fixed-points on Heyting algebras. To this end, we formalize the Intuitionistic Propositional μ-Calculus. An occurrence of a variable x is positive in a formula-term φ if, in the syntax tree of φ, the path from the root to the leaf labeled by this variable contains an even number of nodes labeled by subformulas ψ1 → ψ2 immediately followed by a node labeled by the subformula ψ1 . If, on this path the number of those nodes is odd, then we say that this occurrence of x is negative in φ. A variable x is positive in a formula φ if each occurrence of x is positive in φ. A variable x is negative in a formula φ if each occurrence of x is negative in φ. If we add to the previous grammar (1) the following productions: φ ⇒ μx .φ ,

φ ⇒ νx .φ ,

subject to the restriction that x is positive in φ, we obtain then a grammar for the formulas of IPCμ , the Intuitionistic Propositional μ-Calculus. The semantics of these formulas is the expected one. Let φ be a formula of IPCμ , and let x be positive in φ. Let us denote by Xφ the set of variables having an occurrence in φ. If v : Xφ \ { x } −−→ H is a valuation of all the variables of φ but x in a complete Heyting algebra, then the function φv , defined by h → φv,h/x , is monotone, so μx .φ (resp., νx .φ) is to be evaluated over the least fixed-point (resp., the greatest fixed-point) of this function. A sequent calculus for IPCμ is presented in [7, Sect. 2]. Let us say that a formula φ of IPCμ is fixed-point free if it is a formula of IPC, that is, it does not contain either of the symbols μ, ν.

Fixed-Point Elimination in the Intuitionistic Propositional Calculus

131

Proposition 1. Every formula φ of IPCμ is equivalent to a fixed-point free formula φ . Proof. Clearly, the statement holds if we can show that it holds whenever φ = μx .ψ or φ = νx .ψ, where ψ is a fixed-point free formula. For a natural number n ≥ 0, let ψ n (x) denote the formula obtained by substituting x for ψ n times. Ruitenburg [23] proves that, for each intuitionisitic formula ψ, there exists a number n ≥ 0 such that the formulas ψ n (x) and ψ n+2 (x) are equivalent. If x is positive in ψ, then instantiating x with ⊥, leads to the equivalence ψ n+1 (⊥) ≡ ψ n (⊥), exhibiting ψ n (⊥) as the least fixed-point of ψ. Similarly, ψ n () is the greatest fixed-point of ψ.   While it is an obvious step to derive the previous Proposition from Ruitenburg’s result, there has been no attempt (as far as we know) to compute an upper bound on n ≥ 0 such that ψ n (x) and ψ n+2 (x) are equivalent. Nor is such an n necessarily a tight upper bound for convergence of a formula to its least or greatest fixed-point.

4

Strong Monotone Functions and Fixed-Points

If H is a Heyting algebra and f : H −−→ H is a monotone function, then we say that f is strong if x ∧ f (y) ≤ f (x ∧ y) ,

for any x, y ∈ H.

The interplay between fixed-points and this class of functions has already been emphasized, mainly in the context of categorical proof-theory and semantics of functional programming languages with inductive data types [7, 8]. Lemma 4. A monotone f : H −−→ H is strong if and only if any of the following equivalent conditions holds in H: x ∧ f (y) ≤ f (x ∧ y) , f (x → y) ≤ x → f (y) , x → y ≤ f (x) → f (y) .

(3) (4) (5)

The proof of these equivalences is usual in categorical algebra [17] and therefore it is omitted here. Definition 2. Let H be a Heyting algebra. We say that a function f : H −−→ H is monotone polynomial if there exist a formula φ of the IPC, a variable x positive in φ, and a valuation v : Xφ \ { x } −−→ H such that, for each h ∈ H, we have f (h) = φv,h/x . Proposition 2. Every monotone polynomial f on a Heyting algebra is strong. Proof. Recall that the replacement Lemma holds in the IPC: z ↔ w LJ φ(z) ↔ φ(w). Substituting x for z and x ∧ y for w, and considering that x → y LJ x ↔

132

S. Ghilardi et al.

(x ∧ y), we derive that x → y LJ φ(x) ↔ φ(x ∧ y). Assuming that u is positive in φ(u), we have φ(x) ↔ φ(x ∧ y) LJ φ(x) → φ(x ∧ y) LJ φ(x) → φ(y), whence x → y LJ φ(x) → φ(y). The last relation immediately implies that Eq. (4) from Lemma 4 holds, when f is a monotone polynomial.   It can be shown that the relation f(x) ∧ y = f(x ∧ y) ∧ y holds (for any x, y and) for any polynomial on a Heyting algebra. The analogous remark for Boolean algebras is credited to Peirce, in view of the iteration rule for existential graphs of type Alpha, see [11]. Proposition 3. If f is a strong monotone function on H and a ∈ H, then μ.a → f := a → μ.f ,

μ.a ∧ f := a ∧ μ.f .

(6)

Proof. Let us argue first that first equation holds. To this end, let us set fa (x) =def a → f(x). From f ≤ fa we have Prefa ⊆ Pref . Thus, if p ∈ Prefa , then μx .f(x) = f(μx .f(x)) ≤ f(p) and a → μ.f ≤ a → f(p) = fa (p) ≤ p. That is, a → μ.f is below any element of Prefa . To obtain the proposition, we need to argue that a → μ.f belongs to Prefa . To this end, we notice that { a → p | p ∈ Pref } ⊆ Prefa , since if f(p) ≤ p, then fa (a → p) = a → f(a → p) ≤ a → f(p) ≤ a → p, where we used the fact that f is strong, thus (4) holds. Let us come now to the second equation, for which we set fa (x) =def a ∧ f(x). Suppose a ∧ f(p) ≤ p, so f(p) ≤ a → p. Then f(a → p) ≤ a → f(p) ≤ a → p, using (4), whence μ.f ≤ a → p and a ∧ μ.f ≤ p. Thus we are left to argue that a ∧ μ.f is a prefixed-point of fa . Yet, this is true for an arbitrary prefixed-point p of f: a ∧ f(a ∧ p) ≤ a ∧ f(p) ≤ a ∧ p.   Corollary 1. For each n ≥ 1 and each collection fi , i = 1, . . . , n of monotone polynomials, we have the following distribution law:   fi (x) := μx .fi (x) . (7) μx . i=1,...,n

i=1,...,n

Proof. For n = 1 there is nothing to prove. We suppose therefore that the statement holds for every collection of size n ≥ 1, and prove it holds for a collection of size n + 1. We have   fi (x)) := μx .μy .(fn+1 (y) ∧ fi (x)), by (Diag), μx .(fn+1 (x) ∧ i=1,...,n

i=1,...,n



:= μx .((μy .fn+1 (y)) ∧

fi (x)),

by (6),

i=1,...,n

:= (μy .fn+1 (y)) ∧ μx .(



fi (x)), again by (6),

i=1,...,n

:= (μy .fn+1 (y)) ∧



i=1,...,n

μx .fi (x),

by the IH.  

Fixed-Point Elimination in the Intuitionistic Propositional Calculus

133

The elimination of greatest fixed-points is easy for strong monotone functions (we are thankful to the referee for pointing out the following fact, which greatly simplifies our original argument): Proposition 4. If f : L −−→ L is any strong monotone function on a bounded lattice L, then f2 () = f(). Thus f() is the greatest fixed-point of f. Proof. Indeed, we have f() = f() ∧ f() ≤ f(f() ∧ ) = f2 ().

5

 

A Digression on Fixpoints and Bisimulation Quantifiers

The connection between extremal fixed-points and bisimulation quantifiers, as emphasized in [9], was a main motivation to tackle this research. Although in the end our computations are independant on that, we nevertheless want to have a closer look to the topic (the content of this section is not needed afterwards). It was discovered in [22] that IPC has the uniform interpolation property. As made clear from the title of [22], this property amounts to an internal existential and universal quantification. This result was further refined in [15] to show that any morphism between finitely generated Heyting algebras has a left and a right adjoint. We shall be interested in Heyting algebras H[x] of polynomials with coefficients from H, and to (the left and right adjoints to) the inclusion of H into H[x]. The algebra of polynomials H[x] is formally defined as the coproduct (in the category of Heyting algebras) of H with the free Heyting algebra on one generator. The universal property gives that if h0 ∈ H, then there exists a unique morphism ·h0 /x : H[x] −−→ H such that xh0 /x = h0 and hh0 /x = h, for each h ∈ H. Thus, for f ∈ H[x] and h ∈ H, we can define f(h) = fh/x . It follows from [15] that if H is finitely generated, then the inclusion ix : H −−→ H[x] has both adjoints ∃x , ∀x : H[x] −−→ H, with ∃x  ix  ∀x . In particular, we shall use the unit relation for ∃x : f ≤ ix (∃x (f)) ,

for all f ∈ H[x] .

Identifying h ∈ H with ix (h) ∈ H[x], we can read the above inequality as f ≤ ∃x .f. We can identify a monotone polynomial, as defined in Definition 2, as an element f ∈ H[x] such that fh0 /x ≤ fh1 /x whenever h0 ≤ h1 . Proposition 5. If f is a monotone polynomial on a finitely generated Heyting algebra, then ν.f := ∃x .(x ∧ (x → f(x))) .

(8)

Proof. By the unit relation x ∧ x → f(x) ≤ ∃x .(x ∧ x → f(x)). Recall that evaluation at p ∈ H is a Heyting algebra morphism, thus it is monotone. Therefore, if p ∈ H is a postfixed-point of f, then by evaluating the previous inequality at p, we have p = p ∧ p → f(p) ≤ ∃x .(x ∧ x → f(x)) ,

134

S. Ghilardi et al.

so that ∃x .(x ∧ x → f(x)) is greater than any postfixed-point of f. Let us show that ∃x .(x ∧ x → f(x)) is also a postfixed-point. To this end, it will be enough to argue that x ∧ x → f(x) ≤ f(∃x .(x ∧ x → f(x))) in H[x]. We compute as follows: x ∧ x → f(x) ≤ f(x) ∧ x → f(x) ≤ f(x ∧ x → f(x)), ≤ f(∃x .(x ∧ x → f(x))),

since f is strong, by (3), since f is monotone.  

In a similar fashion, we can prove that if f is a monotone polynomial on a finitely generated Heyting algebra, then μ.f := ∀x .((f(x) → x) → x). As an application, we give an alternative proof of Proposition 4: Corollary 2. If f is a monotone polynomial on a Heyting algebra H, then ν.f := f() .

(9)

Proof. It is easy to see that if f is a monotone polynomial on a finitely generated Heyting algebra, then ∃x .f = f(). Thus we have ν.f = ∃x .(x ∧ (x → f(x))) = ∃x .(x ∧ f(x)) =  ∧ f() = f() . Therefore, if φ is a formula-term whose variables are among set x, y1 , . . . , yn , then the equation φ2 () = φ() holds in the free Heyting algebra on the set { y1 , . . . , yn }. Consequently, the equation f() = f2 () holds in H, making f() into the greatest fixed-point of f.  

6

The Elimination Procedure

In this Section we present our main result, a procedure that both axiomatizes and eliminates least fixed-points of the form μx .φ(x) with φ fixed-point free. Together with the axiomatization of greatest fixed-points given in Sect. 5, the procedure can be extended to a procedure to construct a fixed-point free formula ψ equivalent to a given formula χ of the IPCμ . Definition 3. An occurrence of the variable x is strongly positive in a formulaterm φ if there is no subformula ψ of φ of the form ψ0 → ψ1 such that x is located in ψ0 . A formula-term φ is strongly positive in the variable x if every occurrence of x is strongly positive in φ. An occurrence of a variable x is weakly negative in a formula-term φ if it is not strongly positive. A formula-term φ is weakly negative in the variable x if every occurrence of x is weakly negative in φ. Observe that a variable might be neither strongly positive nor weakly negative in a formula-term. A second key concept for the elimination is the following notion of disjunctive formula. Definition 4. The set of formula-terms that are disjunctive in the variable x is generated by the following grammar: φ ⇒ x | β ∨ φ | φ ∨ β | α → φ | φ ∨ φ,

(10)

where α and β are formulas with no occurrence of the variable x. A formula-term φ is in normal form (w.r.t. x) if it is a conjunction of formula-terms φi , i ∈ I, so that each φi either does not contain the variable x, or it is disjunctive in x.

Fixed-Point Elimination in the Intuitionistic Propositional Calculus

135

Notice that disjunctive formula-terms are strongly positive in x. Due to Eq. (2) and since the usual distributive laws hold in Heyting algebras, we have the following Lemma. Lemma 5. Every strongly positive formula-term is equivalent to a formula-term in normal form. In order to compute the least fixed-point μx .φ, we take the following steps: 1. We rename all the weakly negative occurrences of x in φ to a fresh variable y, so φ(x) = ψ(x, x/y) with ψ strongly positive in x and weakly negative in y. 2. We computea normal form of ψ(x, y), so this formula is equivalent to a conjunction i∈I ψi (x, y) with each ψi disjunctive in x or not containing the variable x. 3. Strongly positive elimination. For each i ∈ I: if x has an occurrence in ψi , we compute then a formula ψi equivalent to the least fixed-point μx .ψi (x, y) and we let ψi = ψi . observe that ψi is weakly negative in y; otherwise,   4. Weakly negative elimination. The formula i∈I ψi (y) is weakly negative in y; we compute a formula χ equivalent to μy . i ψi (y) and return it. The correction of the procedure relies on the following chain of equivalences:  μx .φ(x) = μy .μx .ψ(x, y) = μy .μx . ψi (x, y), where we use (Diag) = μy .

 i∈I

6.1

i∈I

μx .ψi (x, y) = μy .



ψi (y) = χ,

where we have used (7).

i∈I

Strongly positive Elimination

We tackle here the problem of computing the least fixed-point μx .φ of a formulaterm φ which is disjunctive in x. Recall that the formulas α and β appearing in a parse tree as leaves—according to the grammar (10)—do not contain the variable x. We call such a formula α a head subformula of φ, and such a β a side subformula of φ, and thus we put: Head(φ) =def { α | α is a head subformula of φ } , Side(φ) =def { β | β is a side subformula of φ } . Recall that a monotone function f : P −−→ P is inflating if x ≤ f (x). Lemma 6. The interpretation of a strongly positive disjunctive formula φ as a function of x is inflating. The key observation needed to prove Proposition 6 is the following Lemma on monotone inflating functions. In the statement of the Lemma we assume that P is a join-semilattice, and that f ∨ g is the pointwise join of the two functions f and g.

136

S. Ghilardi et al.

Lemma 7. If f, g : P −−→ P are monotone inflating functions, then Pref ∨g = Pref ◦g . Consequently, for any monotone function h : P −−→ P , we have μ.( f ∨ g ∨ h ) :=: μ.( (f ◦ g) ∨ h ) .

(11)

Proof. Observe firstly that Pref ∨g = Pref ∩ Preg . If p ∈ Pref ◦g , then f (p) ≤ f (g(p)) ≤ p and g(p) ≤ f (g(p)) ≤ p, showing that p ∈ Pref ∨g . Conversely, if p ∈ Pref ∨g , then p is a fixed point of both f and g, since these functions are inflating. It follows that f (g(p)) = f (p) = p, showing p ∈ Pref ◦g . We have argued that Pref ∨g coincides with Pref ◦g ; this implies that Pre(f ◦g)∨h = Pref ∨g∨h and, from this equality, Eq. (11) immediately follows.   To ease reading of the next Proposition and of its proof, let us put [α] φ =def α → φ . Proposition 6. If φ is a disjunctive formula-term, then ⎡ ⎤   μ.φ = ⎣ α⎦ ( β) . α∈Head(φ)

(12)

β∈Side(φ)

Proof. For ψ, χ formula-terms, let us write ψ ∼ χ when μ.ψ = μ.χ. We say that a disjunctive formula ψ is reduced (w.r.t. φ) if either it is x, or it is of the form β ∨ x (or x ∨ β) for some β ∈ Side(φ), or of the form [α] x for some α ∈ Head(φ). A set Φ of disjunctive formulas is reduced if every formula in Φ is reduced. We shall compute a reduced set of disjunctive formulas Φk such that φ ∼ Φk . Thus let Φ0 = { φ }. If Φi is not reduced, then there is φ0 ∈ Φi which is not reduced, thus of the form (a) β ∨ ψ (or ψ ∨ β) with ψ = x, or (b) [α] ψ with ψ = x, or (c) ψ1 ∨ ψ2 . According to the case ( ), with ∈ { a, b, c }, we let Φi+1 be (Φi \ { φ0 }) ∪ Ψ where Ψ is as follows: Ψb = { [α] x, ψ }, Ψc = { ψ1 , ψ2 } . Ψa = { β ∨ x, ψ }, By Lemma 7, we have Φi ∼ Φi+1 . Morever, for some k ≥ 0, Φk is reduced and Φk ⊆ { [α] x | α ∈ Head(φ) } ∪ { β ∨ x | β ∈ Side(φ) } ∪ { x }. Consequently    μx .φ(x) = μx . Φk ≤ μx .(x ∨ [α] x ∨ β ∨ x) . (13) α∈Head(φ)

β∈Side(φ)

On the other hand, if α ∈ Head(φ), then φ(x) = ψ1 (x, [α] ψ2 (x)) for some disjunctive formulas ψ1 and ψ2 , so [α] x ≤ [α] ψ2 (x) ≤ ψ1 (x, [α] ψ2 (x)) = φ(x) and, similarly, β ∨ x ≤ φ(x), whenever β ∈ Side(φ). It follows that   x∨ [α] x ∨ β ∨ x ≤ φ(x) , α∈Head(φ)

β∈Side(φ)

Fixed-Point Elimination in the Intuitionistic Propositional Calculus

137

whence, by taking the least fixed-point in both sides of the above inequality, we derive equality in (13). Finally, in order to obtain (12), we compute as follows:   [α] x ∨ β ∨ x) μx .(x ∨ α∈Head(φ) β∈Side(φ)  = μx .([α1 ] . . . [αn ] x ∨ (x ∨ β ∨ x)) β∈Side(φ)







= μx .(⎣

by Lemma 7, with Head(φ) = { α1 , . . . , αn },

α∈Head(φ)



⎡ =⎣ ⎡ =⎣



α⎦ (x ∨



α⎦ μx .(x ∨

α∈Head(φ)



β ∨ x)),



x, α i i=1,...,n

by Lemma 7,

β∈Side(φ)



β ∨ x),

by Proposition 3,

β∈Side(φ)

⎤ α⎦ (

α∈Head(φ)

6.2

since [α1 ] . . . [αn ] x =

α∈Head(φ)



β ∨ x)),

β∈Side(φ)





= μx .(⎣



α⎦ x ∨ (x ∨



β) .

β∈Side(φ)

 

Weakly Negative Elimination

If φ is weakly negative in x then we can write φ(x) = ψ0 (ψ1 (x), . . . , ψn (x)) ,

(14)

for formula-terms ψ0 (y1 , . . . , yn ) and ψi (x), i = 1, . . . , n, such that: (a) all the variables yi are negative in ψ0 ; (b) for i = 1, . . . , n, x is negative ψi . Proposition 7. Let ν1 , . . . , νn be a collection of formula-terms denoting the greatest solution of the system of equations { yi = ψi (ψ0 (y1 , . . . , yn )) | i = 1, . . . , n }. Then ψ0 (ν1 , . . . , νn ) is a formula equivalent to μx .φ(x). Proof. Let v : X \ { x, y1 , . . . , yn } −−→ H be a partial valuation into an Heyting algebra H, put f0 = ψ0 v and, for i = 1, . . . , n, fi = ψi v . Then f0 is a monotone function from [H op ]n to H. Here H op is the poset with the same elements as H but with the opposite ordering relation. Similarly, for 1 ≤ i ≤ n, fi : H −−→ H op . n n If we let ¯f = fi | i = 1, . . . , n ◦ f0 , then ¯f : [H op ] −−→ [H op ] . We exploit next the fact that (·)op is a functor, so that f op : P op −−→ Qop is the same monotone function as f , but considered as having distinct domain and codomain. Then, using (Roll), we can write

138

S. Ghilardi et al.

μ.( f0 ◦ fi | i = 1, . . . , n ) = f0 ( fi | i = 1, . . . , n ◦ f0 ) = f0 ( μ.¯f ) = f0 ( ν.f¯op ) ,

(15)

since the least fixed-point of f in P op is the greatest fixed-point of f op in P . That is, if we consider the function fi | i = 1, . . . , n ◦ f0 as sending a tuple of elements of H (as opposite to H op ) to another such a tuple, then Eq. (15) proves that a formula denoting the least fixed-point of φ is constructible out of formulas for the greatest solution of the system mentioned in the statement of the Proposition.   As far as computing the greatest solution of the system mentioned in the Proposition, this can be achieved by using the Bekic elimination principle, see Lemma 3. This principle implies that solutions of systems can be constructed from solutions of linear systems, i.e. from usual parametrized fixed-points. In our case, as witnessed by Eq. (9), these parametrized greatest fixed-points are computed by substituting  for the fixed-point variable. In the next Section we shall give a more explicit description, by means of approximants, of the least fixed-point of a weakly negative formula φ.

7

Upper Bounds on Closure Ordinals

Recall that Ruitenburg’s result [23] implies that a monotone formula converges to its (parametrized) least fixed-point by iterating the formula n times from ⊥, for some n ≥ 0. That is, we can always substitute μx .φ(x) for some equivalent φn (⊥). We show, in this Section, how to extract, from the procedure just seen, upper bounds for such a number n. Proposition 8. If φ is a disjunctive formula and n is the cardinality of the set Head(φ), then μx .φ(x) = φn+1 (⊥) .

(16)

Proof. We have seen, in the proof of Propositon 6, that [α] x ≤ φ(x) for any α ∈ Head(φ) and, similarly, β ∨ x ≤ φ(x) for any β ∈ Side(φ). Thus we have   β= β ∨ ⊥ ≤ φ(⊥) . β∈Side(φ)

β∈Side(φ)

Let Head(φ) = { α1 , . . . , αn }. Supposing that [αi ] . . . [α1 ] ( β∈Side(φ) β) ≤ φi+1 (⊥), then  [αi+1 ] [αi ] . . . [α1 ] ( β) ≤ [αi+1 ] (φi+1 (⊥)) ≤ φ(φi+1 (⊥)) = φi+2 (⊥) . β∈Side(φ)

Whence



μx .φ(x) = ⎣



i=1,...,n

⎤ αi ⎦ (



β∈Side(φ)

β) = [αn ] . . . [α1 ] (



β) ≤ φn+1 (⊥) .

β∈Side(φ)

 

Fixed-Point Elimination in the Intuitionistic Propositional Calculus

139

The upper bound given in (16) is optimal: if we let φn (x) =def b ∨ i=1,...,n ai → x and consider the Heyting algebra of downsets of P ({ 1, . . . , n }), ⊆ , then, interpreting b as { ∅ } and ai as { s ⊆ { 1, . . . , n } | i ∈ s }, φn converges exactly after n + 1 steps. In order to tackle convergence of weakly negative formulas, we mention some general statements, where we assume that all the posets have a least element. Lemma 8. Convergence for Roll). Let f : P −−→ Q and g : Q −−→ P be monotone functions. If μ.(f ◦ g) = (f ◦ g)n (⊥), then μ.(g ◦ f ) = (g ◦ f )n+1 (⊥). Lemma 9. Convergence for Diag. Let f : P ×P −−→ P be a monotone function. For each p ∈ P , put gp (y) = f (p, x) and h(x) = μy .gx (y). Suppose that, for each p ∈ P , h(p) = μy .f (p, y) = gpn (⊥) and that μx .h(x) = hm (⊥). Then μx .f (x, x) = f nm (⊥, ⊥). For our purposes, the following Lemma provides more accurate upper bounds than Lemma 9. Lemma 10. Let f, g : H −−→ H be strong monotone mappings. If μ.f = f n (⊥) and μ.g = g m (⊥), then μ.f ∧ g = (f ∧ g)n+m−1 (⊥). For the Bekic property we have a similar statement, bounding convergence of f, g by (n + 1)(m + 1) − 1, with m and n being bounds on convergence of μy .g(x, y) and μx .f (x, μy .g(x, y)), respectively. While in general this bound is optimal, the relevant observation is, for our purposes, the following Lemma. Lemma 11. Let {xi = fi (x1 , . . . , xk ) | i = 1, . . . , k} be a monotone system of equations P on some poset with least element ⊥. Suppose that all the functions generated under substitution from { f1 , . . . , fk } ∪ { ⊥ } converge to their parametrized least fixed-point in one step. Then the least solution of this system of equations is obtained by iterating k times f1 , . . . , fk from (⊥, . . . , ⊥) ∈ P k . Proposition 9. Let φ(x) be a weakly negative formula, so that we have a decomposition of the form (14). Then φ(x) converges at its least fixed-point in at most n + 1 steps. Proof. Applying Lemma 11, we have ν.( ψi | i = 1, . . . , n ◦ ψ0 ) = ( ψi | i = 1, . . . , n ◦ ψ0 )n () .

(17)

Considering that μ.φ = μ.(ψ0 ◦ ψi | i = 1, . . . , n ) = ψ0 (ν.( ψi | i = 1, . . . , n ◦ ψ0 )) we can use (17) and Lemma 8 to deduce that μ.φ = (ψ0 ◦ ψi | i = 1, . . . , n )n+1 (⊥) .

 

It is possible to combine Propositions 8 and 9 with Lemma 8 to obtain upper bounds for all formulas. Yet, mainly due to the exponential blow-up in computing

140

S. Ghilardi et al.

an equivalent normal-form of a given formula, that is, step 2 of the procedure described in Sect. 6, these bounds turn out to be exponential functions of the size of the formula. It is possible on the other hand to pinpoint fragments of the IPCμ for which we still have polynomial bounds. For example, if we define a formula-term to be weakly disjunctive if it is generated by the grammar (10), with the difference that we allow x to have weakly negative occurrences in α and β, then bounds are polynomials of order 2.

8

Conclusions

As mentioned in the Introduction, a main motivation for the research described in this paper was to provide in-depth answers to the question of why alternationdepth hierarchies in μ-calculi collapse or are trivial. Until now, the authors dealt with trivial alternation-depth hierarchies. The tools and ideas so far developed still need to be tested when a hierarchy does not completely collapse at its lowest level. In particular, and given the closeness of Intuitionistic Logic with Modal Logic based on transitive frames, it becomes appealing to investigate further connections with existing work on the subject [1, 2, 10, 24]. Compared to other works, such as [20,21], we definitely took an algebraic and constructive approach to the problem of showing definability of least fixedpoints within the IPC. Witnessing the fruitfulness of our approach, the algebra made the goal of computing upper bounds of closure ordinals of the monotone functions denoted by intuitionisitc formulas an accessible task. Let us notice on the way that our work leads to an obvious decision procedure, based on any decision procedure for IPC, for the Intuitionistic Propositional μ-Calculus. This logic, already studied on the side of proof theory and of game semantics [7], should also be of interest in verification, for example when transition systems come with some ordering and upward or downward closed properties are defined by μ-formulas, see [5]. Overall, we believe that understanding extremal fixed-points and more in general fixed-points in an intuitionisitc setting—where sparse but surprising results are known, see for [4] example—is still in quest for an elementary but solid theory to be developed. The present paper is a contribution toward this goal.

References 1. Alberucci, L., Facchini, A.: The modal µ-calculus hierarchy on restricted classes of transition systems. J. Symb. Log. 74(4), 1367–1400 (2009) 2. Alberucci, L., Facchini, A.: On modal µ-calculus and G¨ odel-L¨ ob logic. Studia Logica 91(2), 145–169 (2009) 3. Arnold, A., Niwi´ nski, D.: Rudiments of µ-calculus. Elsevier, Amsterdam (2001) 4. Bauer, A., Lumsdaine, P.L.: On the Bourbaki-Witt principle in toposes. Math. Proc. Cambridge Philos. Soc. 155, 87–99 (2013) 5. Bertrand, N., Schnoebelen, P.: Computable fixpoints in well-structured symbolic model checking. Formal Meth. Syst. Des. 43(2), 233–267 (2013)

Fixed-Point Elimination in the Intuitionistic Propositional Calculus

141

6. Bezhanishvili, N., de Jongh, D.: Intuitionistic Logic. Technical report, Institute for Logic, Language and Computation, Universiteit van Amsterdam, p. 25 (2006) 7. Clairambault, P.: Strong functors and interleaving fixpoints in game semantics. RAIRO - Theory Inf. Appl. 47(1), 25–68 (2013) 8. Cockett, J.R.B., Spencer, D.: Strong categorical datatypes II: a term logic for categorical programming. Theory Comput. Sci. 139(1&2), 69–113 (1995) 9. D’Agostino, G., Hollenberg, M.: Logical questions concerning the mu-calculus: interpolation. Lyndon Los-Tarski. J. Symb. Log. 65(1), 310–332 (2000) 10. D’Agostino, G., Lenzi, G.: On the µ-calculus over transitive and finite transitive frames. Theory Comput. Sci. 411(50), 4273–4290 (2010) 11. Dau, F.: Some notes on proofs with alpha graphs. In: Sch¨ arfe, H., Hitzler, P., Øhrstrøm, P. (eds.) ICCS 2006. LNCS (LNAI), vol. 4068, pp. 172–188. Springer, Heidelberg (2006) ofner, 12. Frittella, S., Santocanale, L.: Fixed-point theory in the varieties Dn . In: H¨ P., Jipsen, P., Kahl, W., M¨ uller, M.E. (eds.) RAMiCS 2014. LNCS, vol. 8428, pp. 446–462. Springer, Heidelberg (2014) 13. Gentzen, G.: Untersuchungen u ¨ ber das logische Schließen. I. Mathematische Zeitschrift 39(1), 176–210 (1935) 14. Ghilardi, S., Zawadowski, M.: Sheaves, Games, and Model Completions: A Categorical Approach to Nonclassical Propositional Logics, 1st edn. Springer Publishing Company, Heidelberg (2011) 15. Ghilardi, S., Zawadowski, M.W.: Model completions, r-Heyting categories. Ann. Pure Appl. Log. 88(1), 27–46 (1997) 16. Kelly, G.: Basic Concepts of Enriched Category Theory. Lecture Notes in Mathematics, vol. 64. Cambridge University Press, Cambridge (1982). Republished on In: Reprints in Theory and Applications of Categories, vol. 10, pp. 1–13 (2005) 17. Kock, A.: Strong functors and monoidal monads. Archiv der Mathematik XXIII 23(1), 113–120 (1972) 18. Kozen, D.: Results on the propositional mu-calculus. Theory Comput. Sci. 27, 333–354 (1983) 19. Lehtinen, K., Quickert, S.: Deciding the first levels of the modal mu alternation hierarchy by formula construction. In: Kreutzer, S. (ed.) 24th EACSL Annual Conference on Computer Science Logic, CSL September 7–10, 2015, Berlin, Germany, vol. 41, pp. 457–471. LIPIcs, Schloss Dagstuhl - Leibniz-Zentrum fuer Informatik (2015) 20. Mardaev, S.I.: Least fixed points in Grzegorczyk’s Logic and in the intuitionistic propositional logic. Algebra Log. 32(5), 279–288 (1993) 21. Mardaev, S.I.: Convergence of positive schemes in our and S4 and Int. Algebra Log. 33(2), 95–101 (1994) 22. Pitts, A.M.: On an interpretation of second order quantification in first order intuitionistic propositional logic. J. Symb. Log. 57(1), 33–52 (1992) 23. Ruitenburg, W.: On the period of sequences (an (p)) in intuitionistic propositional calculus. J. Symb. Log. 49(3), 892–899 (1984) 24. Visser, Albert: L¨ ob’s logic meets the µ-calculus. In: Middeldorp, Aart, Oostrom, Vincent, Raamsdonk, Femke, Vrijer, Roel (eds.) Processes, Terms and Cycles: Steps on the Road to Infinity. LNCS, vol. 3838, pp. 14–25. Springer, Heidelberg (2005)

Verification and Program Analysis

A Theory of Monitors (Extended Abstract) Adrian Francalanza(B) CS, ICT, University of Malta, Msida, Malta [email protected] Abstract. We develop a behavioural theory for monitors — software entities that passively analyse the runtime behaviour of systems so as to infer properties about them. First, we extend the monitor language and instrumentation relation of [17] to handle piCalculus process monitoring. We then identify contextual behavioural preorders that allow us to relate monitors according to criteria defined over monitored executions of piCalculus processes. Subsequently, we develop alternative monitor preorders that are more tractable, and prove full-abstraction for the latter alternative preorders with respect to the contextual preorders.

1

Introduction

Monitors (execution montors [32]) are software entities that are instrumented to execute along side a program so as determine properties about it, inferred from the runtime analysis of the exhibited (program) execution; this basic monitor form is occasionally termed (sequence) recognisers [28]. In other settings, monitors go further and either adapt aspects of the monitored program [7, 11, 22] or enforce predefined properties by modifying the observable behaviour [4, 15, 28]. Monitors are central to software engineering techniques such as monitor-oriented programming [16] and fail-fast design patterns [8] used in fault-tolerant systems [19,34]; they are also used exten sively in runtime verification [27], a lightweight verification technique that attempts to mitigate state explosion problems associated with full-blown verification methods such as model checking. Monitoring setups typically consist of three components: apart from the program being monitored, P , there is the monitor itself, M , and the instrumentation, the mechanism composing the monitor with the program, P  M . The latter gives rise to a software composition relation that has seldom been studied in its own right. This paper investigates compositional reasoning techniques for monitors performing detections (recognisers), composed using the instrumentation relation employed in [17], for programs expressed as piCalculus processes [29, 31], a well-studied concurrency model. We set out to develop monitor preorders M1  M2

(1)

The research was supported by the UoM research fund CPSRP05-04 and the research grant COST-STSM-ECOST-STSM-IC1201-170214-038253. c Springer-Verlag Berlin Heidelberg 2016  B. Jacobs and C. L¨ oding (Eds.): FOSSACS 2016, LNCS 9634, pp. 145–161, 2016. DOI: 10.1007/978-3-662-49630-5 9

146

A. Francalanza

stating that, when instrumented in the context of an arbitrary process P , if P  M1 exhibits certain properties, then P  M2 exhibits them as well. Within this setup, we formalise the possible instrumentation properties one may require from an instrumented (monitored) process, and show how these give rise to different monitoring preorders. Example 1. Consider the monitors M1 and M2 below. M1 monitors for output actions on channel a with a payload that is not in the set C. Apart from detecting the same outputs on channel a, M2 also detects outputs with payload b on channels that are not in D (the + construct acts as an external choice [23]). M1 = match a!x.if x ∈ C then  M2 = (match a!x.if x ∈ C then ) + ( match y!b.if y ∈ D then ) One can argue that M2 is related to M1 , i.e., M1  M2 , since all the detections raised by M1 are also raised by M2 . However, under different criteria, the two monitors would not be related. Consider the case where a ∈ D (or b ∈ C): for a process P exhibiting action a!b, monitor M2 may non-deterministically fail to detect this behaviour; by contrast, M1 always detects the behaviour a!b and, in this sense, M2 does not preserve all the properties of M1 . Monitors are also expected to interfere minimally with the execution of the analysed process, giving rise to other criteria for relating monitors, as we will see in the sequel.  There are various reasons why such preorders are useful. For a start, they act as notions of refinement: they allow us to formally specify properties that are expected of a monitor M by expressing them in terms of a monitor description, SpecM , and then requiring that SpecM  M holds. Moreover, our preorders provide a formal understanding for when it is valid to substitute one monitor implementation for another while preserving elected monitoring properties. We consider a general model that allows monitors to behave non-deterministically; this permits us to study the cases where non-determinism is either tolerated or considered erroneous. Indeed, there are settings where determinism is unattainable (e.g., distributed monitoring [18, 33]). Occasionally, non-determinism is also used to expresses under-specification in program refinement. Although formal and intuitive, the preorders alluded to in (1) turn out to be hard to establish. One of the principal obstacles is the universal quantification over all possible processes for which the monitoring properties should hold. We therefore develop alternative characterisations for these preorders, M1 ≤ M2 , that do not rely on this universal quantification over process instrumentation. We show that such relations are sound wrt. the former monitor preorders, which serves as a semantic justification for the alternative monitor preorders. More importantly, however, it also allows us to use the more tractable alternative relations as a proof technique for establishing inequalities in the original preorders. We also show that these characterisations are complete, thereby obtaining fullabstraction for these alternative preorders. The rest of the paper is structured as follows. Section 2 briefly overviews our process model whereas Sect. 3 introduces our monitor language together with the instrumentation relation. In Sect. 4 we formalise our monitor preorder relations

A Theory of Monitors

147

Syntax P, Q ∈ Proc ::= u!v.P | nil

(output)

| u?x.P

(input)

(nil)

| if u = v then P else Q

(conditional)

| rec X.P

(recursion)

|X

(process var.)

|P Q

(parallel)

| new c.P

(scoping)

Semantics pOut

pThn

pRec

pIn

c!d

!d.P − −− →P

τ

rec X.P − − → P [rec X.P/X] c!d

pCom

τ

if c = d then P else Q − − →Q μ

−−→ P μ Q −−→ P

pPar

−−−→ Q τ

Q− − →P

Q

pRes

−−→ P

−−−→ Q τ

Q− − → new d.(P

Q)

μ

new d.P −−→ new d.P

c?d

− −− →P

Q

μ

c?d

− −− →P

c!d

pCls

pEls

τ

if c = c then P else Q − − →P

c?d

?x.P −−−→ P [d/x]

c!d

pOpn

− −− →P c!d

new d.P − −− →P

Fig. 1. piCalculus syntax and semantics

wrt. this instrumentation. We develop our alternative preorders in Sect. 5, where we also establish the correspondence with the other preorders. Section 6 concludes.

2

The Language

Figure 1 presents our process language, a standard version of the piCalculus. It has the usual constructs and assumes separate denumerable sets for channel names c, d, a, b ∈ Chans, variables x, y, z ∈ Vars and process variables, X, Y ∈ PVars, and lets identifiers u, v range over the sets, Chans ∪ Vars. The input construct, c?x.P , the recursion construct, rec X.P , and the scoping construct, new c.P , are binders where the free occurrences of the variable x, the process variable X, and the channel c resp., are bound in the guarded body P . We write fv(P ), fV(P ), fn(P ), bV(P ), bv(P ) and bn(P ) for the resp. free/bound variables, process variables and names in P . We use standard syntactic conventions e.g., we identify processes up to renaming of bound names and variables (alpha conversion). For arbitrary syntactic objects o, o , we write o  o when the free names in o and o are disjoint e.g., P  Q means fn(P )∩fn(Q) = ∅. The operational semantics of the language is defined by the Labelled Transition System (LTS) shown in Fig. 1. LTS judgements are of the form

148

A. Francalanza μ

I  P −−→ P  where I ⊆ Chans denotes an interface of names known (or shared) by both the process and an implicit observer (with which interactions occur), P is a closed term, and fn(P ) ⊆ I. We write I, c as a shorthand for I ∪ {c} where c ∈ I, and generally assume a version of the Barendregt convention whereby bn(P )  I.1 For arbitrary names c, d, actions μ ∈ Actτ range over input actions, c?d, output actions, c!d, and a distinguished silent action, τ (α ∈ Act ranges over external actions, that exclude τ ). The rules in Fig. 1 are fairly standard, using I for book-keeping purposes relating to free/bound names (we elide symmetric rules for pPar, pCom and pCls); implicitly, c, d ∈ I in rule pOut and c ∈ I in rule pIn (but d is not necessarily in I). We use s, t ∈ Act∗ to denote traces of external actions. Although actions do not include explicit information relating to extruded names, this may be retrieved using I as shown in Definition 1; the absence of action name binding (as in [24, 31]) simplifies subsequent handling of traces. The evolution of I after a transition is determined exclusively by the resp. action of the transition, defined as aftr(I, μ) in Definition 1; note that both the process (through outputs) and the implicit observer (through inputs) may extend I. Definition 1 (Extruded Names and Interface Evolution) def

ext(I, c!d) = {d}\I

def

aftr(I, c!d) = I ∪ {d}

ext(I, τ ) = ∅ aftr(I, τ ) = I

def def

def

ext(I, c?d) = ∅ def

aftr(I, c?d) = I ∪ {d}

We lift the functions in Definition 1 to traces, e.g., aftr(I, s), in the obvious μ1 μ2 way and denote successive transitions IP −−→ P1 and aftr(I, μ1 )P1 −−→ P2 as μ2 μ1 μ μ  −→ to denote  ∃P  ·I P −−→ P  I P −−→ aftr(I, μ1 )P1 −−→ P2 . We write I P − μ μ μ s 1 2 n and I  P =⇒ Q to denote I0  P0 −−→ I1  P1 −−→ I2  P2 . . . −−→ Pn where P0 = P , Pn = Q, I0 = I, Ii = aftr(Ii−1 , μi ) for i ∈ 1..n, and s is equal to μ1 . . . μn after filtering τ labels. Example 2. Consider Psv = rec X.c?x.new d.x!d.X, modelling the idiomatic server that repeatedly waits for requests on c and answers back on the inputted channel with a fresh channel. We can derive the following behaviour wrt. I = {c}: c?a

a!d

I  Psv ==⇒ I, a  new d.a!d.Psv ==⇒ c?a

a!d

I, a, d  Psv ==⇒ I, a, d  new d.a!d.Psv ===⇒ I, a, d, d  Psv Above, bound outputs/inputs [31] are manifested as interface extensions.

3



Monitor Instrumentation

Monitors, M, N ∈ Mon, are syntactically defined by the grammar of Fig. 2. They may reach either of two verdicts, namely detection, , or termination, end, denoting an inconclusive verdict. Our setting is a mild generalisation to that in [17] 1

The rules in Fig. 1 still check explicitly for this; see rules pRes, pCls and pOpn.

A Theory of Monitors

149

Syntax p, q ∈ Pat ::= u?v

(input pattern)

| u!v

(output pattern)

w ∈ Verd ::= end

(termination)

|

(detection)

M, N ∈ Mon ::= w

(verdict)

| p.M

(pattern match)

|M +N

(choice)

| if u = v then M else N

(branch)

| rec X.M

(recursion)

|X

(monitor var.)

Monitor Semantics mVer

α

w −−→ w

mRec

mThn

match(p, α) = σ

mPat

mChL

α

p.M −−→ M σ

mChR

τ

rec X.M − − → M [rec X.M/X] mEls

τ

if c = c then M else N − − →M

μ

M −−→ M μ M + N −−→ M μ

N −−→ N μ M + N −−→ N τ

if c = d then M else N − − →N

Instrumented System Semantics α

−−→ P

iMon α

iTer

−−→ P

α

M −−→ M −−→ P

iAsyP

α

α

M −−→ α

−−→ P

τ

M − − → end

iAsyM

τ

− − → P τ − − → P τ

M − − → M τ − − →

Fig. 2. Monitor syntax, semantics and Instrumentation Semantics

since monitors need to reason about communicated names so as to adequately monitor for piCalculus processes. They are thus equipped with a pattern matching construct (used to observe external actions) and a name-comparison branching construct. The remaining constructs, i.e., external branching and recursion, are standard. Note that, whereas the syntax allows for monitors with free occurrences of monitor variables, monitors are always closed wrt. (value) variables, whereby the outermost occurrence of a variable acts as a binder. E.g., in the monitor (x?c.x!y.if y = d then end else ) pattern x?c binds variable x in the continuation whereas pattern x!y binds variable y. The monitor semantics is defined in terms of an LTS (Fig. 2), modeling the analysis of the visible runtime execution of a process. Following [15, 17, 30], in rule mVer verdicts are able to analyse any external action but transition to the same verdict, i.e., verdicts are irrevocable. By contrast, pattern-guarded monitors only transition when the action matches the pattern, binding pattern variables to the resp. action names, match(p, α) = σ, and substituting them in the continuation, M σ; see rule mPat. The remaining transitions are unremarkable.

150

A. Francalanza

A monitored system, P  M , consists of a process, P , instrumented with a monitor, M , analysing its (external) behaviour. Figure 2 defines the instrumentation semantics for configurations, I  P  M , i.e., systems augmented with an interface I, where again we assume P is closed and fn(P ) ⊆ I. The LTS semantics follows [7, 17] and relies on the resp. process and monitor semantics of Figs. 1 and 2. In rule iMon, if the process exhibits the external action α wrt. I, and the monitor can analyse this action, they transition in lock-step in the instrumented system while exhibiting same action. If, however, a process exhibits an action that the monitor cannot analyse, the action is manifested at system level while the monitor is terminated ; see rule iTer. Finally, iAsyP and iAsyM allow monitors and processes to transition independently wrt. internal moves, i.e., our instrumentation forces process-monitor synchronisation for external actions only, which constitute our monitorable actions. We note that, as is expected of recognisers, the process drives the behaviour of a monitored system: if the process cannot α-transition, the monitored system cannot α-transition either. Example 3. Recall Psv from Example 2. Using the semantics of Fig. 2, one can derive the monitored execution leading to a detection   below, when composed with the monitor M1 = c?y.y!z.if z = c then end else  , subject to I  = {c, a}: c?a · a!d

I   Psv  M1 =====⇒ Psv   Contrastingly, for the same I  , monitoring with M2 = (y!z.if y = a then end else ) does not lead to a detection for the same trace, because the first action, c?a (an input) cannot pattern match with the output pattern, y!z. In fact, rule iTer terminates the monitor after transition c?a, so as to avoid erroneous detections. τ

I   Psv  M2 −→ c?x.new d.x!d.Psv  M2 a!d

c?a

−−−→ new d.a!d.Psv  end ==⇒ Psv  end

(2)

For illustrative purposes, consider N = c?y.if y = c then end else y!z.c?z., another monitor. We have the following (dissected) transition sequence for I = {c}: I  Psv  N

  τ c?a − → · −−→ I, a  new d.a!d.Psv  if a = c then end else a!z.c?z. (3)   τ (4) −→ I, a  new d.a!d.Psv  a!z.c?z. a!d

−−−→ τ

(I, a, d)  Psv  c?d. c?b

− → · −−→ 

b!d

−−−→





(5)

(I, a, d, b)  new d .b!d .Psv  end

(6)

(I, a, d, b, d )  Psv  end

(7)

In (3) the server (asynchronously) unfolds (pRec and iAsyP) and inputs on c the fresh name a (pIn); the monitor can analyse c?a (mPat where match(c?y, c?a) = {y → a}), transitioning accordingly (iMon) while learning the fresh name a

A Theory of Monitors

151

for future monitoring. At this stage, the instrumentation temporarily stalls the process, even though it can produce the (scope extruding) output a!d. More precisely, although the monitor cannot presently analyse a!d, the rule iTer — which terminates the monitor execution — cannot be applied, since the monitor can silently transition and potentially become capable of analysing the action. This is indeed the case, (4) using mEls, resulting in the second monitoring transition, (5) using iMon, where the monitor learns the second fresh name d, this time originating from the monitored process. After another unfolding, the process is ready to input on c again. However, the monitor cannot match c?b (match(c?d, c?b) is undefined) and since the monitor cannot silently transition either, it is terminated (iTer) while still allowing the process to proceed, (6). In (7), verdicts allow monitored processes to execute without any hindrance.  Example 3 highlights two conflicting instrumentation requirements. On the one hand, monitors should interfere minimally with the execution of a monitored process where, observationally, a monitored process should behave like the original one. On the other hand, instrumentation must also ensure bona fide detections, e.g., in (2) and (6), terminating monitoring when the observed process behaviour does not correspond, (through rule iTer). But in order to do this while avoiding premature termination, instrumentation needs to allow for monitor internal computation, e.g., (4). Unfortunately, the premise caveat τ M−  → in rule iTer — necessary to prevent premature terminations — allows monitors to affect (indirectly) process behaviour. For instance the monitor Ω below: Ω = rec X.(if c = c then X else X) τ

τ

Ω  = if c = c then Ω else Ω

τ

(8) α

is divergent, i.e., Ω − → Ω − → Ω − → . . ., and unresponsive, i.e., ∀α · Ω −  −→  α  −→. As a result, it suppresses every process external behaviour when and Ω − α α  −→ instrumented: for arbitrary I  P we can show I  P  Ω −  −→ and I  P  Ω  − for any α, since rules iMon and iTer cannot be applied. We revisit this point in Sect. 4. We conclude with the property stating that verdicts are irrevocable. s

Theorem 1 (Definite Verdicts). I  P  w =⇒ Q  M implies M = w

4

Monitor Preorders

We can use the formal setting presented in Sect. 3 to develop the monitor preorders discussed in the Introduction. We start by defining the monitoring predicates we expect to be preserved by the preorder; a number of these predicates rely on computations and detected computations, defined below. Definition 2 (Detected Computations). The transition sequence s

τ

τ

τ

I  P  M =⇒ I0  P0  M0 −→ I1  P1  M1 −→ I2  P2  M2 −→ . . . is called an s-computation if it is maximal (i.e., either it is infinite or it is finite and cannot be extended further using τ -transitions). An s-computation is called  detected (or a detected computation along s) iff ∃n ∈ N · Mn = .

152

A. Francalanza

One criteria for comparing monitors considers the verdicts reached after observing a specific execution trace produced by the process under scrutiny. The semantics of Sect. 3 assigns a passive role to monitors, prohibiting them from influencing the branching execution of the monitored process. Definition 2 thus differentiates between detected computations, identifying them by the visible trace that is dictated by the process (over which the monitor should not have any control). Example 4. Consider P = new d.(d!  d?.c!a  d?.c!b), I = {c, b, a} and monitors: M1 = c!a. + c!b. M4 = c!a. + c!b. + c!b.end

M2 = c!a. + c!b.end M3 = c!a. M5 = c!a. + c!b. + c!a.end + c!b.end

Configurations I  Mi  P for i ∈ 1..5 exhibit detecting computations along s = c!a. . For trace t = c!b. , configurations I  Mj  P for j ∈ {1, 4, 5} detect t-computations as well, whereas the resp. configurations with M2 and M3 do not. Although the configuration with M1 always detects along t, those with M4 and M5 may fail to detect it along such a trace. Similarly, configuration with M1 and M4 deterministically detect along trace s, but I  M5  P does not.  Example 4 suggests two types of computation detections that a monitor may exhibit. Definition 3 (Potential and Deterministic Detection). M potentially detects for I  P along trace s, denoted as pd(M, I, P, s), iff there exists a detecting s-computation from I  P  M . M deterministically detects for I  P along trace s, denoted as dd(M, I, P, s), iff all s-computation from I  P  M are detecting.  s

Remark 1. If a monitored process cannot produce trace s, i.e., I  P =  ⇒, then pd(M, I, P, s) is trivially false and dd(M, I, P, s) is trivially true for any M .  The detection predicates of Definition 3 induce the following monitor preorders (and equivalences), based on the resp. detection capabilities. Definition 4 (Potential and Deterministic Detection Preorders) M pd N

=

def

∀I, P, s · pd(M, I, P, s) implies pd(N, I, P, s)

M dd N

def

∀I, P, s · dd(M, I, P, s) implies dd(N, I, P, s)

=

M∼ =pd N and M ∼ =dd N are the kernel equivalences induced by the resp. preorders, def i.e., M ∼ =pd N = (M pd N and N pd M ), and similarly for M ∼ =dd N . We write  M pd N in lieu of (M pd N and N pd M ) and similarly for M dd N . Example 5. Recall the monitors defined in Example 4. It turns out that M2 M5

∼ =pd dd

M3 M2

pd ∼ =dd

M5 M3

∼ =pd ∼ =dd

M4 M4

∼ =pd dd

M1 M1

(9) (10)

Note that, whereas M5 can potentially detect more computations that M2 and M3 , (9), it can deterministically detect less computations than these monitors  (10); in fact, M5 cannot deterministically detect any computation.

A Theory of Monitors

153

As opposed to prior work on monitors [2, 10, 15], the detection predicates in Definition 3 consider monitor behaviour within an instrumented system. Apart from acting as a continuation for the study in [17], this setup also enables us to formally justify subtle monitor orderings, Example 6, and analyse peculiarities brought about by the instrumentation relation, Example 7. Example 6. Using the shorthand τ.M for rec X.M where X ∈ fV(M ), we have: ∼ =pd τ.

but

 dd τ.

For process P = Ω, defined as in (8), predicate dd(, I, P, ) holds trivially, but predicate dd(τ., I, P, ) does not, due of the non-detecting -computation τ τ τ τ →·− → Ω  τ. − →·− → . . ., refuting the inequality  dd τ..  I  Ω  τ. − Example 7. Recalling the divergent monitor Ω from (8), we have:   end ∼ dd c!a. dd Ω dd  =dd c!a.end   ∼ Ω =dd Ω +  ∼ Ω + end dd =dd rec X.(τ.X + ) but Ω + c!a. ∼ Ω + c!a. pd Ω +  =dd Ω + 

(11) (12) (13)

In (11), every computation starting with monitor  is trivially detected. Monitor Ω limits all computations to -computations, i.e., irrespective of I  P , configuration I  P  Ω exhibits no s-computations for any s where |s| > 0, rendering dd(Ω, I, P, s) for |s| > 0 vacuously true (see Remark 1). By contrast, s dd(c!a., I, P, s) holds only whenever s = c!a.t (for arbitrary t) and I  P =⇒. In (12), monitor Ω + end does not deterministically detect any computation: when composed with an arbitrary I  P , it clearly can never reach a detection, but it can neither prohibit the process from producing visible actions, as in the case of Ω (see rules mVer, mChR and iMon). Monitor Ω +  can either behave like Ω or transition to  after one external action observed; in both cases, it deterministically detects all s-computation where |s| > 0. The monitor rec X.(τ.X + ) first silently transitions to (τ.rec X.(τ.X + )) +  and then either transitions back to the original state or else transitions to  with an external action; in either case, when composed with any process I  P , it also deterministically detects all s-computation for |s| > 0. In (13), although monitor Ω + c!a. potentially detects less computations than Ω +  (e.g., for I = {c, a, b}, P = c!b.nil and s = c!b. , the predicate pd(Ω+, I, P, s) holds but pd(Ω+c!a., I, P, s) does not), both deterministically detect the same computations, i.e., all s-computation where |s| > 0. Specifically, if a process being monitored, say I  P , can produce an action other than c!a, the instrumentation with monitor Ω + c!a. restrains such an action, since the monitor cannot transition with that external action (it can only transition with c!a) but, at the same time, it can τ -transition (see rules iMon and iTrm).  The preorders in Definition 4 are not as discriminating as one might expect.

154

A. Francalanza

Example 8. Consider the monitor Many = x?y. + x!y.. Many ∼ =pd Many + Ω ∼ c!a.end =pd c!a.end + c!a.Ω

and and

Many ∼ =dd Many + Ω c!a.end ∼ =dd c!a.end + c!a.Ω

(14) (15)

Intuitively, Many potentially and deterministically detects any s-computation when |s| > 0. It turns out that Many + Ω produces the same potential and deterministic detections, yielding the resp. equalities in (14). In (15), neither monitor produces any potential or deterministic detections and they are thus equivalent according to the resp. kernel equivalences of Definition 4.  There are however settings where the equalities established in Example 8 are deemed too coarse. E.g., in (15), whereas monitor c!a.end is innocuous when instrumented with a process, monitor c!a.end+c!a.Ω may potentially change the observed behaviour of the process under scrutiny after the action c!a is emitted (by suppressing external actions, as explained in Example 7); a similar argument applies for the monitors in (14). We thus define a third monitor predicate called transparency [5,15, 28], stating that whenever a monitored process cannot perform an external action, it must be because the (unmonitored) process is unable to perform that action (i.e., the monitoring does not prohibit that action). Definition 5 (Transparency Preorder). M is transparent for I P wrt. trace  s α   ⇒ s, denoted as tr(M, P, I, s), iff I P  M =⇒ Q N and aftr(I, s)  (Q  N ) = α implies aftr(I, s)  Q =  ⇒. We define the induced preorder as expected: M tr N

def

∀I, P, s · tr(M, I, P, s) implies tr(N, I, P, s)

=



Although the preorders in Definitions 4 and 5 are interesting in their own right, we define the following relation as the finest monitor preorder in this paper. Definition 6 (Monitor Preorder) M N

def

=

M pd N and M dd N and M tr N



Example 9. We have Many  Many + Ω because Many tr Many + Ω, since ¬tr(Many + Ω, I, P, s) for I = {c, a}, P = c!a.c!a.nil and s = c!a. . Similarly, we also have c!a.end  c!a.end + c!a.Ω.  Inequalities from the preorders of Definitions 4 and 5 are relatively easy to repudiate. For instance, we can use P, I and t from Example 4 as counter examples to show that pd(M5 , I, P, t) and ¬pd(M3 , I, P, t), thus disproving M5 pd M3 . However, it is much harder to show that an inequality from these preorders holds because we need to consider monitor behaviour wrt. all possible processes, interfaces and traces. As shown in Examples 6, 7 and 8, this often requires intricate reasoning in terms of the three LTSs defined in Figs. 1 and 2.

A Theory of Monitors

5

155

Characterisation

We define alternative monitor preorders for which positive statements about their inequalities are easier to establish. The new preorders are defined exclusively in terms of the monitor operational semantics of Fig. 2, as opposed to how they are affected by arbitrary processes as in Definition 6 (which considers also the process and instrumentation LTSs). We show that the new preorders coincide with those in Sect. 4. Apart from equipping us with an easier mechanism for determining the inequalities of Sect. 4, the correspondence results provide further insight into the properties satisfied by the preorders of Definitions 4 and 5. We start with the potential-detection preorder. We first define a restricted α monitor LTS that disallows idempotent transitions from verdicts, w −−→ w: these are redundant when considering the monitor operational semantics in isolation. α Note, however, that we can still used rule mVer, e.g., to derive  + M −−→r . Definition 7 (Restricted Monitor Semantics). A derived monitor transiμ μ tion, M −−→r N , is the least relation satisfying the conditions M −−→ N and s  M = w. M =⇒r N denotes a transition sequence in the restricted LTS. We use the restricted LTS to limit the detecting transition sequences on the left of the implication of Definition 8. However, we permit these transitions to be matched by transition sequences in the original monitor LTS, so as to allow the monitor to the right of the inequality to match the sequence with a prefix of α visible actions (which can then be padded by  −−→  transitions as required). Definition 8 (Alternative Potential Detection Preorder) def

s

M pd N = ∀s · M =⇒r 

implies

s

N =⇒ 

Theorem 2 (Potential-Detection Preorders). M pd N

iff

M pd N

Example 10. By virtue of Theorem 2, to show that Ω + c!a. pd Ω +  from c!a

(13) of Example 7 holds, we only need to consider Ω + c!a. ==⇒r , which can c!a

be matched by Ω + ==⇒ . Similarly, to show (x!a.if x = c then  else end) pd c!a

, we only need to consider (x!a.if x = c then  else end) ==⇒r , matched by c!a

 −−→ .



For the remaining characterisations, we require two divergence judgements. Definition 9 (Divergence and Strong Divergence) – M ↑ denotes that M diverges, meaning that it can produce an infinite transiτ τ τ tion sequence of τ -actions M −→ M  −→ M  −→ . . . – M ⇑ denotes that M strongly diverges, meaning that it cannot produce finite τ τ τ  →.  transition sequence of τ -actions M −→ M  −→ . . . M  − τ

Lemma 1. M −→

implies

M = w

156

A. Francalanza

The alternative preorder for deterministic detections, Definition 11 below, is based on three predicates describing the behaviour of a monitor M along a trace s. The predicate blk(M, s) describes the potential for M to block before it can complete trace s. Predicate fl(M, s) describes the potential for failing after monitoring trace s, i.e., an s-derivative of M reaches a non-detecting state from which no further τ actions are possible, or it diverges (implicitly, by Lemma 1, this also implies that the monitors along the diverging sequences are never detecting). Finally nd(M, s) states the existence of a non-detecting s-derivative of M . Definition 10 (Monitor Block, Fail and Non-Detecting) s

def

τ

α

1 N−  → and N − → blk(M, s) = ∃s1 , α, s2 , N · s = s1 αs2 and M ==⇒   def s τ fl(M, s) = ∃N · M =⇒ N and (N =  and N −  →) or N ↑ def

s

nd(M, s) = ∃N · M =⇒ N and N =  Note that blk(M, s) implicitly requires |s| ≥ 1 for the predicate to hold. Corollary 1. blk(M, s)

implies



∀t · blk(M, st)

Definition 11 (Alternative Deterministic Detection Preorder) ⎧ ⎪ ⎨blk(N, s) implies blk(M, s) or fl(M, s) def M dd N = ∀s · fl(N, s) implies blk(M, s) or fl(M, s) ⎪ ⎩ nd(N, s) implies nd(M, s) or blk(M, s) We write M dd N to denote the kernel equality (M dd N and N dd M ).  Theorem 3 (Deterministic-Detection Preorders). M dd N iff M dd N Example 11. Consider M = c?a.end + x!b.end and N = c?a.end. By virtue of Theorem 3, to determine M dd N we prove M dd N as follows: 1. We have blk(N, s) whenever s = αs and α = c?a. We have two subcases: – If match(x!b, α) is undefined, we show blk(M, αs ) by first showing that blk(M, α ) and then generalising the result for arbitrary s using Corollary 1. – If ∃σ · match(x!b, α) = σ, we show fl(M, αs ) by first showing fl(M, α ) and then generalising the result for arbitrary s using Theorem 1. 2. For any s, we have fl(N, c?a.s): we can show fl(M, c?a.s), again using Theorem 1 to alleviate the proof burden. 3. For any s, we have nd(N, c?a.s): the required proof is analogous to the previous case.  Example 12. Due to full abstraction (i.e., completeness), we can alternatively disprove  dd τ. from Example 6 by showing that  dd τ.: we can readily argue that whereas nd(τ., ), we cannot show either nd(, ) or blk(, ).  Example 13. Recall the equalities Ω ∼ =dd rec X.(τ.X + ) claimed in =dd Ω +  ∼ (12) of Example 7. It is arguably easier to determine these equalities by considering only the monitor LTSs to show that Ω dd Ω +  dd rec X.(τ.X + ) since:

A Theory of Monitors

157

1. For any s = we have ¬blk(Ω, s), ¬blk(Ω +, s) and ¬blk(rec X.(τ.X +), s). 2. We only have fl(Ω, ), fl(Ω + , ) and fl(rec X.(τ.X + ), ). 3. Similarly, we only have nd(Ω, ), nd(Ω + , ) and nd(rec X.(τ.X + ), ).  Remark 2. The alternative preorder in Definition 11 can be optimised further using refined versions of the predicates fl(M, s) and nd(M, s) that are defined in terms of the restricted monitor transitions of Definition 7, as in the case of Definition 3.  The alternative transparency preorder, Definition 13 below, is defined in terms of divergence refusals which, in turn, rely on strong divergences from Definition 9. Intuitively, divergence refusals are the set of actions that cannot be performed whenever a monitor reaches a strongly divergent state following the analysis of trace s. These actions turn out to be those that are suppressed on a process when instrumented with the resp. monitor. Definition 12 (Divergence Refusals)  def s α dref(M, s) = α | ∃N · M =⇒ N and N ⇑ and N =⇒ Definition 13 (Alternative Transparency Preorder) def

M tr N = ∀s · dref(N, s) ⊆ dref(M, s) Theorem 4 (Transparency Preorders). M tr N

iff

M tr N

Example 14. Recall monitors c!a.end and c!a.end+c!a.Ω from (15) of Example 8. The inequality c!a.end + c!a.Ω tr c!a.end follows trivially from Theorem 4, since ∀s · dref(c!a.end, s) = ∅. The symmetric case, c!a.end tr c!a.end + c!a.Ω, can also be readily repudiated by plying Theorem 4. Since dref((c!a.end + c!a.Ω), c!a. ) = Act (and dref(c!a.end, c!a. ) = ∅) we trivially obtain a violation of the set inclusion requirements of Definition 13.  Example 15. Recall again Ω +  and rec X.(τ.X + ) from (12). We can differentiate between these monitors from a transparency perspective, and Theorem 4 permits us to do this with relative ease. In fact, whereas dref((Ω + ), ) = Act τ (since Ω+ −→ Ω and dref(Ω, ) = Act) we have dref((rec X.(τ.X +)), ) = ∅; for all other traces |s| ≥ 1 we obtain empty divergence refusal sets for both monitors. We thus can positively conclude that Ω +  tr rec X.(τ.X + ) while  refuting rec X.(τ.X + ) tr Ω + . Definition 14 (Alternative Monitor Preorder) M N

def

=

M pd N and M dd N and M tr N

Theorem 5 (Full Abstraction). M  N

iff

M N



158

6

A. Francalanza

Conclusion

We have presented a theory for (recogniser) monitors that allows us to substitute a monitor M1 in a monitored process P  M1 by another monitor M2 while guaranteeing the preservation of a number of monitoring properties relating to (behaviour) detection and monitor interference. The theory is compositional, since it enables us to ensure the preservation of properties by analysing the resp. monitors M1 and M2 in isolation, without needing to consider the process being monitored, P (which may be arbitrarily complex). To the best of our knowledge, it is the first monitor theory of its kind and could be used to alleviate efforts for proving monitors correct e.g., [26]. The concrete contributions are: 1. The definition of three monitor preorders, each requiring the preservation of different monitoring properties: Definitions 3, 4 and 5. 2. The characterisation of these preorders in terms of alternative preorders that are more tractable, Theorems 2, 3 and 4. Related and Future Work. The instrumentation relation we consider is adopted from [17] and embodies synchronous instrumentation (where the external actions constituted the monitorable actions). Synchronous instrumentation is the most prevalent method used in monitoring tools (e.g., [1, 9,14, 25]) because it carries benefits such as timely detections. There are however variants such as asynchronous instrumentation (e.g., [12, 20]) as well as hybrid variations (e.g., [6, 7, 9, 30]). Our theory should be applicable, at least in part, to these variants. In runtime verification, three-verdict monitors [2, 10, 15, 17] are often considered, where detections are partitioned into acceptances and rejections. The monitors studied here express generic detections only; they are nevertheless maximally expressive for branching-time properties [17]. They also facilitate comparisons with other linear-time preorders (see below). We also expect our theory to extend smoothly to settings with acceptances and rejections. Our potential and deterministic detection preorders are reminiscent of the classical may and must preorders of [13,23] and, more recently (for the deterministic detection preorder), of the subcontract relations in [3, 21]. However, these relations differ from ours in a number of respects. For starters, the monitor instrumentation relation of Fig. 2 assigns monitors a passive role whereas the parallel composition relation composing processes (servers in [3, 21]) with tests (clients in [3, 21]) invites tests to interact with the process being probed. Another important difference is that testing preorders typically relate processes, whereas our preorders are defined over the adjudicating entities i.e., the monitors. The closest work in this regard is that of [3], where the authors develop a must theory for clients. Still, there are significant discrepancies between this must theory and our deterministic detection preorder (further to the differencies between the detected (monitored) computations of Definition 2 and the successful computations under tests of [3, 13, 23] as outlined above — success in the compliance relation of [21] is even more disparate). Concretely, in our setting we have equalities such as c!a. ∼ =dd c!a. + c!b.end (see (10) of Example 6),

A Theory of Monitors

159

which would not hold in the setting of [3] since their client preorder is sensitive to external choices (dd is not because monitored executions are distinguished by their visible trace). The two relations are in fact incomparable, since divergent processes are bottom elements in the client must preorder of [3], but they are not in dd . In fact, we have Ω dd Ω + end in (12) of Example 7 or, more clearly, Ω dd Ω + α.end; at an intuitive level, this is because the instrumentation relation of Fig. 2 prioritises silent actions over external actions that cannot be matched by the monitor. Transparency is usually a concern for enforcement monitors whereby the visible behaviour of a monitored process should not be modified unless it violates some specified property [5,15, 28]. We adapted this concept to recognisers, whereby the process behaviour should never be suppressed by the monitor. To our knowledge, the only body of work that studies monitoring for the piCalculus is [5,11, 22], and focusses on synthesising adaptation/enforcement monitors from session types. The closest to our work is [5]: their definitions of monitor correctness are however distinctly different (e.g., they are based on branching-time equivalences) and their decomposition methods for decoupling the monitor analysis from that of processes rely on static type-checking. Acknowledgements. The paper benefited from discussions with Luca Aceto, Giovanni Bernardi, Matthew Hennessy and Anna Ing´ olfsd´ ottir.

References 1. Barringer, H., Falcone, Y., Havelund, K., Reger, G., Rydeheard, D.E.: Quantified event automata: Towards expressive and efficient runtime monitors. In: Giannakopoulou, D., M´ery, D. (eds.) FM 2012. LNCS, vol. 7436, pp. 68–84. Springer, Heidelberg (2012) 2. Bauer, A., Leucker, M., Schallhart, C.: Runtime verification for LTL and TLTL. TOSEM 20(4), 14 (2011) 3. Bernardi, G., Hennessy, M.: Mutually testing processes. LMCS 11(2:1), 1–23 (2015) 4. Bielova, N., Massacci, F.: Do you really mean what you actually enforced? Edited automata revisited. Int. J. Inf. Secur. 10(4), 239–254 (2011) 5. Bocchi, L., Chen, T.-C., Demangeon, R., Honda, K., Yoshida, N.: Monitoring networks through multiparty session types. In: Beyer, D., Boreale, M. (eds.) FORTE 2013 and FMOODS 2013. LNCS, vol. 7892, pp. 50–65. Springer, Heidelberg (2013) 6. Cassar, I., Francalanza, A.: On synchronous and asynchronous monitor instrumentation for actor systems. FOCLASA 175, 54–68 (2014) 7. Kane, A., Chowdhury, O., Datta, A., Koopman, P.: A case study on runtime monitoring of an autonomous research vehicle (ARV) system. In: Bartocci, E., et al. (eds.) RV 2015. LNCS, vol. 9333, pp. 102–117. Springer, Heidelberg (2015). doi:10.1007/978-3-319-23820-3 7 8. Cesarini, F., Thompson, S.: Erlang Programming. O’Reilly, Sebastopol (2009) 9. Chen, F., Ro¸su, G.: MOP: An efficient and generic runtime verification framework. In: OOPSLA, pp. 569–588. ACM, (2007) 10. Cini, C., Francalanza, A.: An LTL proof system for runtime verification. In: Baier, C., Tinelli, C. (eds.) TACAS 2015. LNCS, vol. 9035, pp. 581–595. Springer, Heidelberg (2015)

160

A. Francalanza

11. Coppo, M., Dezani-Ciancaglini, M., Venneri, B.: Self-adaptive monitors for multiparty sessions. In: PDP, pp. 688–696. IEEE Computer Society (2014) 12. D’Angelo, B., Sankaranarayanan, S., S´ anchez, C., Robinson, W., Finkbeiner, B., Sipma, H.B., Mehrotra, S., Manna, Z.: LOLA: Runtime monitoring of synchronous systems. In: TIME, IEEE (2005) 13. De Nicola, R., Hennessy, M.C.B.: Testing equivalences for processes. TCS 34(1–2), 83–133 (1984) 14. Decker, N., Leucker, M., Thoma, D.: jUnitRV –adding runtime verification to junit. In: Brat, G., Rungta, N., Venet, A. (eds.) NFM 2013. LNCS, vol. 7871, pp. 459–464. Springer, Heidelberg (2013) 15. Falcone, Y., Fernandez, J.-C., Mounier, L.: What can you verify and enforce at runtime? STTT 14(3), 349–382 (2012) 16. Formal Systems Laboratory. Monitor Oriented Programming. University of Illinois at Urbana Champaign. http://fsl.cs.illinois.edu/index.php/Monitoring-Oriented Programming 17. Kane, A., Chowdhury, O., Datta, A., Koopman, P.: A case study on runtime monitoring of an autonomous research vehicle (ARV) system. In: Bartocci, E., et al. (eds.) RV 2015. LNCS, vol. 9333, pp. 102–117. Springer, Heidelberg (2015). doi:10.1007/978-3-319-23820-3 7 18. Francalanza, A., Gauci, A., Pace, G.J.: Distributed system contract monitoring. JLAP 82(5–7), 186–215 (2013) 19. Francalanza, A., Hennessy, M.: A theory for observational fault tolerance. JLAP 73(1–2), 22–50 (2007) 20. Francalanza, A., Seychell, A.: Synthesising correct concurrent runtime monitors. FMSD 46(3), 226–261 (2015) 21. Castagna, G., Gesbert, N., Padovani, L.: A theory of contracts for web services. ACM Trans. Program. Lang. Syst. 31(5), 1–61 (2009) 22. Giusto, C.D., Perez, J.A.: Disciplined structured communications with disciplined runtime adaptation. Sci. Comput. Program. 97(2), 235–265 (2015) 23. Hennessy, M.: Algebraic Theory of Processes. MIT Press, Cambridge (1988) 24. Hennessy, M.: A Distributed Pi-Calculus. Cambridge University Press, Cambridge (2007) 25. Kim, M., Viswanathan, M., Kannan, S., Lee, I., Sokolsky, O.: Java-MaC: A runtime assurance approach for Java programs. FMSD 24(2), 129–155 (2004) 26. Kane, A., Chowdhury, O., Datta, A., Koopman, P.: A case study on runtime monitoring of an Autonomous Research Vehicle (ARV) system. In: Bartocci, E., et al. (eds.) RV 2015. LNCS, vol. 9333, pp. 102–117. Springer, Heidelberg (2015). doi:10.1007/978-3-319-23820-3 7 27. Leucker, M., Schallhart, C.: A brief account of runtime verification. JLAP 78(5), 293–303 (2009) 28. Ligatti, J., Bauer, L., Walker, D.: Edit automata: enforcement mechanisms for run-time security policies. Int. J. Inf. Secur. 4(1–2), 2–16 (2005) 29. Milner, R.: Communication and Concurrency. Prentice-Hall Inc, Upper Saddle River (1989) 30. Ro¸su, G., Havelund, K.: Rewriting-based techniques for runtime verification. Autom. Softw. Engg. 12(2), 151–197 (2005) 31. Sangiorgi, D., Walker, D.: PI-Calculus: A Theory of Mobile Processes. Cambridge University Press, Cambridge (2001) 32. Schneider, F.B.: Enforceable security policies. ACM Trans. Inf. Syst. Secur. 3(1), 30–50 (2000)

A Theory of Monitors

161

33. Sen, K., Vardhan, A., Agha, G., Rosu, G.: Efficient decentralized monitoring of safety in distributed systems. In: ICSE, pp. 418–427. IEEE (2004) 34. Verissimo, P., Rodrigues, L.: Distributed Systems for System Architects. Kluwer Academic Publishers, Norwell (2001)

Contextual Approximation and Higher-Order Procedures Ranko Lazi´c and Andrzej S. Murawski(B) DIMAP and Department of Computer Science, University of Warwick, Coventry, UK [email protected]

Abstract. We investigate the complexity of deciding contextual approximation (refinement) in finitary Idealized Algol, a prototypical language combining higher-order types with state. Earlier work in the area established the borderline between decidable and undecidable cases, and focussed on the complexity of deciding approximation between terms in normal form. In contrast, in this paper we set out to quantify the impact of locally declared higher-order procedures on the complexity of establishing contextual approximation in the decidable cases. We show that the obvious decision procedure based on exhaustive β-reduction can be beaten. Further, by classifying redexes by levels, we give tight bounds on the complexity of contextual approximation for terms that may contain redexes up to level k, namely, (k−1)-EXPSPACE-completeness. Interestingly, the bound is obtained by selective β-reduction: redexes from level 3 onwards can be reduced without losing optimality, whereas redexes up to order 2 are handled by a dedicated decision procedure based on game semantics and a variant of pushdown automata.

1

Introduction

Contextual approximation (refinement) is a fundamental notion in programming language theory, facilitating arguments about program correctness [14] as well as supporting formal program development [5]. Intuitively, a term M1 is said to contextually approximate another term M2 , if substituting M1 for M2 in any context will not lead to new observable behaviours. Being based on universal quantification over contexts, contextual approximation is difficult to establish directly. In this paper, we consider the problem of contextual approximation in a higher-order setting with state. Contextual reasoning at higher-order types is a recognised challenge and a variety of techniques have been proposed to address it, such as Kripke logical relations [3] or game models [2]. In this work, we aim to understand the impact of locally defined higher-order procedures on the complexity of establishing contextual approximation. Naturally, one would expect the complexity to grow in the presence of procedures and it should grow as the type-theoretic order increases. We shall quantify that impact by providing tight Research supported by EPSRC (EP/J019577/1, EP/M011801/1). c Springer-Verlag Berlin Heidelberg 2016  B. Jacobs and C. L¨ oding (Eds.): FOSSACS 2016, LNCS 9634, pp. 162–179, 2016. DOI: 10.1007/978-3-662-49630-5 10

Contextual Approximation and Higher-Order Procedures

163

complexity bounds for contextual approximation in our higher-order framework. The results demonstrate that, from the complexity-theoretic point of view, it is safe to inline procedures only down to a certain level. Below that level, however, it is possible to exploit compositionality to arrive at better bounds than those implied by full inlining. The vehicle for our study is Idealized Algol [1, 13], the protypical language for investigating the combination of local state with higher-order procedures. In order to avoid obviously undecidable cases, we restrict ourselves to its finitary variant IAf , featuring finite base types and no recursion (looping is allowed, though). Both semantic and syntactic methods were used to reason about Idealized Algol [1,12] in the past. In particular, on the semantic front, there exists a game model that captures contextual approximation (in the sense of inequational full abstraction) via complete-play inclusion. Earlier work in the area [6, 9,11] used this correspondence to map out the borderline between decidable and undecidable cases within IAf . The classification is based on type-theoretic order: a term is of order i if its type is of order at most i and all free variables have order less than i. We write IAi for the set of IAf -terms of order i. It turns out that contextual approximation is decidable for terms of all orders up to 3, but undecidable from order 4 onwards. The work on decidability has also established accurate complexity bounds for reasoning about contextual approximation between terms in β-normal form as well as terms with the simplest possible β-redexes, in which arguments can only be of base type. For order-3 terms, the problem can be shown EXPTIME-complete [11], while at orders 0, 1, 2 it is PSPACE-complete [10]. In this paper, we present a finer analysis of the decidable cases and consider arbitrary β-redexes. In particular, functions can be used as arguments, which corresponds to the inlining of procedures. We evaluate the impact of redexes by introducing a notion of their level: the level of a β-redex (λx.M )N will be the order of the type of λx.M . Accordingly, we can split IAi into sublanguages IAki , in which terms can contain redexes of level up to k. IA0i is then the normal-form case and IA1i is the case of base-type arguments. Obviously, the problem of contextually approximating IAki (i ≤ 3, k ≥ 2) terms can be solved by β-reduction (and an appeal to the results for IA0i ), but this is known to result in a k-fold exponential blow-up, thus implying a (k + 1)EXPTIME upper bound for IAk3 . This bound turns out to be suboptimal. One could lower it by reducing to IA1i instead, which would shave off a single exponential, but this is still insufficient to arrive at the optimal bound. It turns out, however, that reducing IAk3 terms to IA23 (all redexes up to order 3 are eliminated) does not lead to a loss of optimality. To work out the accurate bound for the IA23 case, one cannot apply further β-reductions, though. Instead we devise a dedicated procedure based on game semantics and pushdown automata. More specifically, we introduce a variant of visibly pushdown automata [4] with -transitions and show how to translate IA23 into such automata so that the accepted languages are faithful representations of the term’s game semantics [1]. The translation can be performed in exponential time and, crucially, the automata correspoding to IA23 -terms satisfy a boundedness property: the stack symbols pushed on the stack

164

R. Lazi´c and A.S. Murawski

during -moves can only form contiguous segments of exponential length with respect to the term size. This allows us to solve the corresponding inclusion problem in exponential space with respect to the original term size. Consequently, we can show that IA23 contextual approximation is in EXPSPACE. The above result then implies that program approximation of IAk3 -terms is in (k − 1)-EXPSPACE. Furthermore, we can prove matching lower bounds for IAk1 . The table below summarises the complexity results. The results for k ≥ 2 are new. k=1

k≥2

IAk1 PSPACE-complete [10]

k=0

PSPACE-complete [10]

(k − 1)-EXPSPACE-complete

PSPACE-complete [10]

PSPACE-complete [10]

(k − 1)-EXPSPACE-complete

IAk2 IAk3

2

EXPTIME-complete [11] EXPTIME-complete [11] (k − 1)-EXPSPACE-complete

Idealized Algol

We consider a finitary version IAf of Idealized Algol with active expressions [1]. Its types are generated by the following grammar. θ ::= β | θ → θ

β ::= com | exp | var

IAf can be viewed as a simply-typed λ-calculus over the base types com, exp, var (of commands, expressions and variables respectively) augmented with the constants listed below skip : com i : exp (0 ≤ i ≤ max ) succ : exp → exp pred : exp → exp seqβ : com → β → β deref : var → exp ifzeroβ : exp → β → β → β assign : var → exp → com cellβ : (var → β) → β while : exp → com → com mkvar : (exp → com) → exp → var

where β ranges over base types and exp = { 0, · · · , max }. Other IAf -terms are formed using λ-abstraction and application Γ  M : θ → θ Γ  N : θ Γ  M N : θ

Γ, x : θ  M : θ Γ  λxθ .M : θ → θ using the obvious rules for constants and free identifiers. Each of the constants corresponds to a different programming feature. For instance, the sequential composition of M and N (typically denoted by M ; N ) is expressed as seqβ M N , assignment of N to M (M := N ) is represented by assignM N and cellβ (λx.M ) amounts to creating a local variable x visible in M (new x in M ). mkvar is the so-called bad-variable constructor that makes it possible to construct terms of type var with prescribed read- and write-methods. whileM N corresponds to while M do N . We shall write Ωβ for the divergent constant that can be defined using while 1 do skip. The operational semantics of IAf , based on call-by-name evaluation, can be found in [1]; we will write M ⇓ if M reduces to skip. We study the induced contextual approximation.

Contextual Approximation and Higher-Order Procedures MA×B = MA + MB λA×B = [λA , λB ] A×B = A + B

MA⇒B = MA + MB λA⇒B = [λA , λB ] A⇒B = B + (IB × IA ) + (

A

165

∩ (MA × MA ))

λA reverses the ownership of moves in A while preserving their kind. Fig. 1. Constructions on arenas

Definition 1. We say that Γ  M1 : θ contextually approximates Γ  M2 : θ if, for any context C[−] such that C[M1 ], C[M2 ] are closed terms of type com, we have C[M1 ] ⇓ implies C[M2 ] ⇓. We then write Γ  M1  ∼ M2 . Even though the base types are finite, IAf contextual approximation is not decidable [9]. To obtain decidability one has to restrict the order of types, defined by: ord(β) = 0

ord(θ → θ ) = max(ord(θ) + 1, ord(θ )).

Definition 2. Let i ≥ 0. The fragment IAi of IAf consists of IAf -terms x1 : θ1 , · · · , xn : θn M : θ such that ord(θj ) < i for any j = 1, · · · , n and ord(θ) ≤ i. Contextual approximation is known to be decidable for IA1 , IA2 and IA3 [11], but it is undecidable for IA4 [9]. Definition 3. – The level of a β-redex (λxθ .M )N is the order of the type of λxθ .M . – A term has degree k if all redexes inside it have level at most k. – IAki is the subset of IAi consisting of terms whose degree is at most k. β-reduction can be used to reduce the degree of a term by one at an exponential cost. Lemma 1. Let d ≥ 1. A λ-term M of degree d can be reduced to a term M  of degree d − 1 with a singly-exponential blow-up in size.

3

Games

Game semantics views computation as an exchange of moves between two players, called O and P. It interprets terms as strategies for P in an abstract game derived from the underlying types. The moves available to players as well as the rules of the game are specified by an arena. Definition 4. An arena is a triple A =  MA , λA , A , where – MA is a set of moves; – λA : MA → { O, P } × { Q, A } is a function indicating to which player (O or P ) a move belongs and of what kind it is (question or answer);

166

R. Lazi´c and A.S. Murawski

Acom

Aexp

run

q

done

0

···

Avar read max

0

···

write(0) max

···

write(max)

ok

Fig. 2. Arenas for base types

– A ⊆ (MA + {  }) × MA is the so-called enabling relation, which must satisfy the following conditions: • If  enables a move then it is an O-question without any other enabler. A move like this is called initial and we shall write IA for the set containing all initial moves of A. • If one move enables another then the former must be a question and the two moves must belong to different players. Arenas used to interpret the base types of IAf are shown in Fig. 2: the moves at the bottom are answer-moves. Product and function-space arenas can be constructed as shown in Fig. 1. Given an IAf -type θ, we shall write [[θ]] for the corresponding arena obtained compositionally from Acom , Aexp and Avar using the ⇒ construction. Given arenas, we can play games based on a special kind of sequences of moves. A justified sequence s in an arena A is a sequence of moves in which every move m ∈ IA must have a pointer to an earlier move n in s such that n A m. n is then said to be the justifier of m. It follows that every justified sequence must begin with an O-question. Given a justified sequence s, its O-view s and P-view s are defined as follows, where o and p stand for an O-move and a P-move respectively:  =   = 

so = so so = o (if o is initial)

sp = sp

s o t p = so p s p t o  = sp o .

A justified sequence s satisfies visibility condition if, in any prefix tm of s, if m is a non-initial O-move then its justifier occurs in t and if m is a P-move then its justifier is in t. A justified sequence satisfies the bracketing condition if any answer-move is justified by the latest unanswered question that precedes it. Definition 5. A justified sequence is a play iff O- and P -moves alternate and it satisfies bracketing and visibility. We write PA for the set of plays in an arena A. A play is single-threaded if it contains at most one occurrence of an initial move. The next important definition is that of a strategy. Strategies determine unique responses for P (if any) and do not restrict O-moves. Definition 6. A strategy in an arena A (written as σ : A) is a non-empty prefix-closed subset of single-threaded plays in A such that:

Contextual Approximation and Higher-Order Procedures

167

skip : com run done i : exp qi succ : exp 1 ⇒ exp q q1 max i=0 i1 ((i + 1) mod max ) pred : exp 1 ⇒ exp q q1 max i=0 i1 ((i − 1) mod max ) ifzeroβ : exp 3 ⇒ β 2 ⇒ β 1 ⇒ β max i=1 i3 ) q2 a2 a q β a q q3 03 q1 a1 a + q β a q q3 ( seqβ : com 2 ⇒ β 1 ⇒ β q β a q run 2 done 2 q1 a1 a deref : var 1 ⇒ exp q read 1 max i=0 i1 i assign : var 2 ⇒ exp 1 ⇒ com run q1 max i=0 i1 write(i)2 ok 2 done cellβ : ( var 1,1 ⇒ β 1 ) ⇒ β max ∗ ∗ ∗ i=0 write(i)1,1 ok 1,1 (read 1,1 i1,1 ) ) a1 a q β a qq1 (read 1,1 01,1 ) ( mkvar : ( exp 2,1 ⇒ com 2 ) ⇒ exp 1 ⇒ var max ∗ read q1 ( max i=0 i1 i) + i=0 write(i) run 2 (q2,1 i2,1 ) done 2 ok ∗ while : exp 2 ⇒ com 1 ⇒ com run q2 ( max i run 2 1 done 1 q2 ) 02 done i=1

Fig. 3. Strategies for constants. Only complete plays are specified.

(i) whenever sp1 , sp2 ∈ σ and p1 , p2 are P-moves then p1 = p2 ; (ii) whenever s ∈ σ and so ∈ PA for some O-move o then so ∈ σ. We write comp (σ) for the set of non-empty complete plays in σ, i.e. plays in which all questions have been answered. An IAf term Γ  M : θ, where Γ = x1 : θ1 , · · · , xn : θn , is interpreted by a strategy (denoted by [[Γ  M : θ]]) in the arena [[Γ  θ]] = [[θ1 ]] × · · · × [[θn ]] ⇒ [[θ]]. The denotations are calculated compositionally starting from strategies corresponding to constants and free identifiers [1]. The latter are interpreted by identity strategies that copy moves across from one occurrence of the same arena to the other, subject to the constraint that the interactions must be plays. Strategies corresponding to constants are given in Fig. 3, where the induced complete plays are listed. We use subscripts to indicate the origin of moves. Let σ : A ⇒ B and τ : B ⇒ C. In order to compose the strategies, one first defines an auxiliary set σ † of (not necessarily single-threaded) plays on A ⇒ B that are special interleavings of plays taken from σ (we refer the reader to [1] for details). Then σ; τ : A ⇒ C can be obtained by synchronising σ † and τ on B-moves and erasing them after the synchronisation. The game-semantic interpretation captures contextual approximation as follows. Theorem 1 [1]. Γ  M1  ∼ M2 if and only if comp [[Γ  M1 ]] ⊆ comp [[Γ  M2 ]]. Remark 1. σ † is an interleaving of plays in σ that must itself be a play in PA⇒B . For instance, only O is able to switch between different copies of σ and this can only happen after P plays in B. We shall discuss two important cases in detail, namely, B = [[β]] and B = [[βk → · · · → β1 → β]]. – If B = [[β]] then a new copy of σ can be started only after the previous one is completed. Thus σ † in this case consists of iterated copies of σ.

168

R. Lazi´c and A.S. Murawski

– If B = [[βk → · · · → β1 → β]] then, in addition to the above scenario, a new copy of σ can be started by O each time P plays qi (question from βi ). An old copy of σ can be revisited with ai , which will then answer some unanswered occurrence of qi . However, due to the bracketing condition, this will be possible only after all questions played after that qi have been answered, i.e. when all copies of σ opened after qi are completed. Thus, in this particular case, σ † contains “stacked” copies of σ. Consequently, we can capture X = {  } ∪ comp (σ † ) in this case by equation  X = {ε} ∪ { q A0 qi1 X ai1 A1 . . . qim X aim Am a X | qA0 qi1 ai1 A1 . . . qim aim Am a ∈ comp (σ)} where Aj ’s stand for (possibly empty and possibly different) segments of moves from A. Note that, in a play of σ, qi will always be followed by ai .

4

Upper Bounds

We shall prove that contextual approximation of IA23 terms can be decided in exponential space. Thanks to Lemma 1, this will imply that approximation of IAk3 (k ≥ 2) terms is in (k − 1)-EXPSPACE. In Sect. 5 we will show that these bounds are tight. This shows that by firing redexes of level higher than 3 we do not lose optimal complexity. However, if redexes of order 2 were also executed (i.e. first-order procedures were inlined), the problem would be reduced to IA13 and the implied bound would be 2-EXPTIME, which turns out suboptimal. In what follows, we show that contextual approximation of IA23 terms is in EXPSPACE. To that end, we shall translate the terms to automata that represent their game semantics. The alphabet of the automata will consist of moves in the corresponding games. Recall that each occurrence of a base type in a type contributes distinct moves. In order to represent their origin faithfully, we introduce a labelling scheme based on subscripts. First we discuss how to label occurrences of base types in types. Let Θ be a type of order at most 3. Then Θ ≡ Θm → · · · → Θ1 → β and Θi ’s are of order at most 2. Consequently, for each 1 ≤ i ≤ m, we have Θi ≡ Θi,mi → · · · → Θi,1 → βi and Θi,j ’s are of order at most 1. Thus, we have Θi,j ≡ βi,j,mi,j → · · · → βi,j,1 → βi,j . Note that the above decomposition assigns a sequence of subscripts to each occurrence of a base type in Θ. Observe that ord(Θ) = 3 if and only if some occurrence of a base type gets subscripted with a triple. Next we are going to employ the subscripts to distinguish base types in IA3 typing judgments. Definition 7. A third-order typing template Ψ is a sequence x1 : θ1 , · · · , xn : θn , θ, where ord(θi ) ≤ 2 (1 ≤ i ≤ n) and ord(θ) ≤ 3. To label θ1 , · · · , θn , θ we will use the same labelling scheme as discussed above but, to distinguish θi ’s from θ and from one another, we will additionally use superscripts xi for the former. The labelling scheme will also be used to identify

Contextual Approximation and Higher-Order Procedures

169

moves in the corresponding game. Recall that the game corresponding to a thirdorder typing template will have moves from M[[θ1 ]] +· · ·+M[[θn ]] +M[[θ]] . The superand subscripts will identify their origin in a unique way. Example 1. Let Ψ ≡ x1 : (com → exp) → var, x2 : com → exp → var, ((com → 1 → expx1 1 ) → exp) → var) → com. Here is the labelling scheme for Ψ : x1 : (comx1,1 x2 x2 x1 x2 var , x2 : com2 → exp1 → var , ((com1,1,1 → exp1,1 ) → var1 ) → com. In the 1 , run x2 2 , q1x2 , corresponding games, among others, we will thus have moves run x1,1 x2 read , run 1,1,1 as well as run. Our representation of game semantics will need to account for justification pointers. Due to the well-bracketing condition, pointers from answers need not be represented explicitly. Moreover, because of the visibility condition, in our case x and qi,j,k . Such we only need to represent pointers from moves of the shapes qi,j x pointers must point at some moves of the form qi and qi,j respectively. In order to represent a pointer we are going to place a hat symbol above both the source x  and target of the pointer, i.e. we shall also use “moves”of the form q i,j,k i,j , q x (sources) and qi , q (targets) the latter hatted moves will only be used if the i,j former exist in the sequence. Similarly to [8], we shall represent a single play by several sequences of (possibly hatted) moves under the following conditions: – whenever a target-move of the kind discussed above is played, it may or may not be hatted in the representing sequences of moves, – if a target-move is hatted, all source-moves pointing at the target move are also hatted, – if a target-move is not hatted, no source-moves pointing at the move are hatted. Note that this amounts to representing all pointers for a selection of possible targets, i.e. none, one or more (including all). Because the same -symbol is used to encode each pointer, in a single sequence there may still be ambiguities as to the real target of a pointer. However, among the representing plays we will also have plays representing pointers only to single targets, which suffice to recover pointer-related information. This scheme works correctly because only pointers from P-moves need to be represented and the strategies are deterministic (see the discussion at the end of Sect. 3 in [11]). Example 2. The classic examples of terms that do need explicit pointers are the Kierstaad terms  K1 , K2 : ((com1,1,1 → com1,1 ) → com1 ) → com defined by com .xi )). To represent the corresponding Ki ≡ λf (com→com)→com .f (λxcom 1 .f (λx2 strategies the following sequences of moves will be used (among others). – K1 : q q1 q1,1 q1 q1,1 q1,1,1 (zero targets), q q1 q 1,1 q1 q1,1 q 1,1,1 (one target),   q q1 q1,1 q1 q 1,1 q1,1,1 (one target), q q1 q 1,1 q1 q 1,1 q 1,1,1 (two targets). – K2 : q q1 q1,1 q1 q1,1 q1,1,1 (zero targets), q q1 q 1,1 q1 q1,1 q1,1,1 (one target),   q q1 q1,1 q1 q 1,1 q 1,1,1 (one target), q q1 q 1,1 q1 q 1,1 q 1,1,1 (two targets).

170

R. Lazi´c and A.S. Murawski

To represent strategies corresponding to IA23 -terms we are going to define an extension of visibly pushdown automata [4]. The alphabet will be divided push-, pop- and no-op-letters corresponding to possibly hatted moves. Additionally, we will use -transitions that can modify stack content, albeit using a distinguished stack alphabet. Definition 8. Let Ψ = x1 : θ1 , · · · , xm : θm , θ be a third-order typing template and let M = M[[θ1 ]] +· · · +M[[θn ]] +M[[θ]] . Below we shall refer to the various components of M using subscripts and superscripts according to the labelling scheme introduced earlier, also using q and a for questions and answers respectively. We define the sets Σpush , Σpop , Σnoop as follows. xh  xh xh – Σpush = {qi,j,k , q i,j,k | qi,j,k ∈ M} ∪ {qi,j , qi,j | qi,j ∈ M} xh xh – Σpop = {ai,j,k | ai,j,k ∈ M} ∪ {ai,j | ai,j ∈ M} xh xh  – Σnoop = (M \ (Σpush ∪ Σpop )) ∪ {q i,j | qi,j,k ∈ M} ∪ {qi | qi,j ∈ M}

Σpush and Σpop contain exclusively P- and O-moves respectively, while we can O P , Σnoop for subsets of Σnoop find both kinds of moves in Σnoop . Let us write Σnoop consisting of O- and P-moves respectively. The states of our automata will be partitioned into states at which O is to move (O-states) and whose at which P should reply (P-states). Push-moves and -transitions are only available at P-states, while pop-transitions belong to O-states. No-op transitions may be available from any kind of state. Further, to reflect determinacy of strategies, P-states allow for at most one executable outgoing transition, which may be labelled with an element of Σ P (push or no-op) or be silent (noop, push or pop). Definition 9. Let Ψ be a third-order typing template. A Ψ -automaton A is a tuple (Q, Σ, Υ, δ, i, F ) such that – Q = QO + QP is a finite set of states partitioned into O-states and P-states, – Σ = Σ O + Σ P is the finite transition alphabet obtained from Ψ as above, O and Σ P = Σpush + partitioned into O- and P-letters, where Σ O = Σpop +Σnoop P Σnoop , – Υ = Υ Σ + Υ  is a finite stack alphabet partitioned into Σ-symbols and symbols, O O O + δnoop + δ P is a transition function consisting of δpop : QO × Σpop × – δ = δpop P O O O P P P ΥΣ  Q , δnoop : Q × Σnoop  Q and δ : Q  (Σpush × QO × ΥΣ ) + P (Σnoop × QO ) + QP + (QP × Υ ) + (Υ  QP ), O – i ∈ Q is an initial state, and – F ⊆ QO is a set of final states. Ψ -automata are to be started in the initial state with empty stack. They will accept by final state, but whenever this happens the stack will be empty anyway. Clearly, they are deterministic. The set of words derived from runs will be referred to as the trace-set of A, written T (A). We write L(A) for the subset of T (A) consisting of accepted words only. The Ψ -automata to be constructed will satisfy an additional run-time property called P-liveness: whenever the automaton reaches a configuration (q, γ) ∈ QP × Υ from (i, ), δ P will provide a unique executable transition.

Contextual Approximation and Higher-Order Procedures

171

Remark 2. In what follows we shall reason about IA23 terms by structural induction. The base cases are the constants and identifiers Γ, f : θ  f : θ, where ord(θ) ≤ 2. For inductive cases, we split the rule for application into linear application and contraction. Γ  M : θ → θ Δ  N : θ ord(θ → θ ) ≤ 2 Γ, Δ  M N : θ

Γ, x : θ, y : θ  M : θ Γ, x : θ  M [x/y] : θ

Note that the restriction on θ → θ is consistent with the fact that the level of redexes cannot exceed 2 and free identifiers have types of order at most 2. The relevant λ-abstraction rule is Γ, x : θ  M : θ ord(θ → θ ) ≤ 3. Γ  λxθ .M : θ → θ This stems from the fact that we are considering IA3 . Lemma 2. Let x1 : θ1 , · · · , xm : θm  M : θ be an IA23 -term and let σ = [[x1 : θ1 , · · · , xm : θm  M : θ]]. There exists a P-live (x1 : θ1 , · · · , xm : θm , θ)automaton AM , constructible from M in exponential time, such that T (AM ) and L(AM ) represent respectively σ and comp (σ) (in the sense of our representation scheme). Proof. Translation by structural induction in IA23 . The base cases corresponding to the special constants can be resolved by constructing finite automata, following the description of the plays in Fig. 3. For free identifiers, automata of a similar kind have already been constructed as part of the translation of normal forms in [11]. We revisit them below to show which moves must be marked to represent pointers. Let θ be a second-order type. Then x : θ  x: θ is interpreted by the identity x x strategy, which has complete plays of the form qa qq Xa a, where X is given by the context-free grammar below. When writing qa , we mean summing up over all pairs of moves of the indicated shape available in the associated arena M such that q M a. Below we also use the condition ∃q.qi  q to exclude moves of the form qi that do not enable any other questions (such moves are never targets of justification pointers). X→ Yi →

   | ( qi ai qix qi Yi∗ ai axi ) X | ( qi ai qix qi (Yi )∗ ai axi ) X ∃q.qi q   x x x x  Yi → qi,j ai,j qi,j qi,j Xai,j ai,j qi,j ai,j qi,j qi,j Xai,j ai,j

To capture X, we can construct Ax as in [11], by pushing return addresses when x x x reading qi,j , q i,j and popping them at ai,j . Note that this simply corresponds to interpreting recursion in the grammar. λ-abstraction and contraction are interpreted by renamings of the alphabet, so it remains to consider the hardest case of (linear) application. The rule simply corresponds to composition: in any cartesian-closed category [[Γ, Δ  M N : θ ]] is equal (up to currying) to [[Δ  N : θ]]; [[  λxθ .λΓ.M x : θ → (Γ → θ )]]. Note

172

R. Lazi´c and A.S. Murawski

that in our case ord(θ) ≤ 1, i.e. Remark 1 will apply and the strategy for M N can be obtained by running the strategy for M , which will call copies of N , whose interleavings will obey the stack discipline. To model the interaction, let us consider the moves on which the automata will synchronise. Since ord(θ) ≤ 1, the moves that will interact will be of the form q, a, qi , ai from N ’s point of view and qk , ak , qk,i , ak,i from M ’s viewpoint for some k. Thus, given AM = (QM , ΣM , ΥM , iM , δM , FM ) and AN = (QN , ΣN , ΥN , iN , δN , FN ), we let AM N = (QM N , ΣM N , ΥM N , iM , δM N , FM ), where QM N ΣM N ΣM N ΥM N  ΥM N

= QM + (QO M × QN ) = (ΣM \ {qk , ak , qk,i , ak,i }) + (ΣN \ {q0 , a0 , q1 , a1 }) = ΥM + Υ N  = ΥM + ΥN + QO M

The decomposition of ΣM N into push-, pop- and noop-letters is inherited from the constituent automata. We specify the transition function δM N below using derivation rules referring to transitions in AM and AN . A push-transition reading x/γ

x,γ

 x and pushing γ will be labelled with −−→. Dually, −−→ will represent a pop. x stands for any transition involving x, where x could also be . – AM ’s non-interacting transitions are copied over to AM N . x 

→AM s s−

x ∈ (ΣM \ {qk , ak , qk,i , ak,i }) + {}

x 

s− →AM N s

– M calls N (left) and N returns from the call (right). qk

q

s −→AM s

iN − →AN t



a

k  s −→ AM s

s− →AM N (s , t)

a



t− →AN f ∈ FN

(s , t) − →AM N s

– N ’s non-interacting transitions are copied over while keeping track of AM ’s state. x 

→AN t t− x 

(s, t) − →AM N

s ∈ QO M , x ∈ (ΣN \ {q0 , a0 , q1 , a1 }) ∪ {}

(s, t )

– N calls its argument (left) and the argument returns (right). qk,i

s −−→AM s /t

qi

t −→AN t

(s, t) −−→AM N s

ak,i

s −−→AM s

a

i  t −→ AN t

,t

s −−→AM N (s , t )

Note that the interaction involves moves that are not used to represent pointers, i.e. whenever pointers are represented they remain the same as they were in the original strategies, which is consistent with the definition of composition. O The states in QM N are divided into O- and P -states as follows: QO M N = QM + O O P P O P (QM × QN ) and QM N = QM + (QM × QN ). The correctness of the construction

Contextual Approximation and Higher-Order Procedures

173

follows from the fact that it is a faithful implementation of legal interactions (see, e.g., [7]), as discussed in Remark 1. P-liveness follows from the fact the constituent strategies are P-live and that the construction simulates interaction sequences, including infinite chattering.   Our next step will be to analyse the shape of reachable configurations of AM . We aim to understand how many elements of Υ can occur consecutively on the stack. Definition 10. Suppose (q, γ) ∈ Q × (ΥΣ ∪ Υ )∗ . The -density of γ is defined to be the length of the longest segment in γ consisting solely of consecutive elements from Υ . While the size of stacks corresponding to IA23 terms is unbounded (consider, for instance, x : θ  x : θ with θ = (com → com) → com), -density turns out to be bounded. We shall prove that it is exponential with respect to the size of the original term. This will be crucial to obtaining our upper bound. The main obstacle to proving this fact is the case of composition M N . As discussed in Remark 1, M “stacks up” copies of N and we would first like to obtain a bound on the number of nested calls to N that are not separated by a move from Σpush (such moves block the growth of -density). For this purpose, we go back to plays and analyse sequences in which the relevant questions are pending: a pending question is one that has been played but remains unanswered. Observe that sequences of pending questions are always alternating. We will not be interested in the specific questions but only in their kinds, as specified by the table below. x Question q qi , q x qi,j , qix qi,j,k , qi,j

Kind

01

2

3

Definition 11. Let s be a play. We define pend(s) to be the sequence from {0, 1, 2, 3}∗ obtained from s by restricting it to pending questions and replacing each question with the number corresponding to its kind. Thus, any non-empty even-length play s, pend(s) will match the expression 0(12 + 32)∗ (1 + 3). We say that the (12)-potential of s is equal to k if k is the largest k such that pend(s) = · · · (12)k · · · . In other words, the (12)-potential of a play is the length of the longest segment (12)k in pend(s). Lemma 3. Let Γ  M : θ be an IA23 -term. Then the (12)-potential of any play in [[Γ  M ]] is bounded and the bound bM is exponential in the size of M . Lemma 3 is a key technical result needed to establish the following boundedness property that is satisfied by automata representing IA23 -terms. Lemma 4. Let Γ  M : θ be an IA23 -term and consider AM constructed in Lemma 2. There exists a bound dM , exponential in the size of M , such that the -density of configurations reachable by AM is bounded by dM .

174

R. Lazi´c and A.S. Murawski

Next we derive a bound on plays witnessing failure of contextual approximation in IA23 . Consider IA23 -terms Γ  M1 , M2 : θ and let σi = [[Γ  Mi : θ]] for i = 1, 2. Given a play, let its height be the maximum number of pending questions from Σpush occurring in any of its prefixes. Note that, for plays from σi , this will be exactly the maximum number of symbols from ΥΣ that will appear on the stack of AMi at any point of its computation. Lemma 5. There exists a polynomial p such that if comp σ1 \ comp σ2 is not empty then it contains a play of height p(n1 + n2 ), where n1 , n2 are the numbers of states in AM1 and AM2 respectively. Theorem 2. For IA23 -terms Γ  M1 , M2 : θ, one can decide Γ  M1  ∼ M2 in exponential space. Proof. Note that this boils down to testing emptiness of comp σ1 \ comp σ2 . By Lemma 5, it suffices to guess a play whose height is polynomial in the size of AM1 , AM2 , i.e. exponential with respect to term size. Moreover, by Lemma 4, the -density of the corresponding configurations of AM1 and AM2 will also be exponential. Thus, in order to check whether a candidate s is accepted by AM1 and rejected by AM2 , we will only need to consider stacks of exponential size wrt M1 , M2 . Consequently, the guess can be performed on the fly and verified in exponential space. Because NEXPSPACE=EXPSPACE, the result follows. Corollary 1. For k ≥ 2, contextual approximation of IAk3 -terms is in (k − 1)EXPSPACE.

5

Lower Bounds

Here we show that contextual approximation of IAk1 -terms is (k−1)-EXPSPACEhard for k ≥ 2. Note that this matches the upper bound shown for IAk3 -terms and will allow us to conclude that contextual approximation in IAk1 , IAk2 and IAk3 is (k − 1)-EXPSPACE-complete. Our hardness results will rely on nesting of function calls and iteration afforded by higher-order types. Below we introduce the special types and terms to be used. Let k, n ∈ N. Define the type n by 0 = com and n + 1 = n → n. Note that ord(n) = n. Also, let Exp(k, n) be defined by Exp(0, n) = n and Exp(k + 1, n) = 2Exp(k,n) . Given k ≥ 2, consider the term twicek = λxk−1 .λy k−2 .x(xy) : k. Definition 12. Let k ≥ 2. Writing M n N as shorthand for M (M · · · (M N ) · · · ), 

n

let us define a family of terms {nestn,k } with f : 1, x : 0  nestn,k : 0 by taking nestn,k ≡ (twicenk gk−1 )gk−2 · · · g1 g0 , where g0 ≡ x, g1 ≡ f and gi ≡ twicei for i > 1. The terms have several desirable properties, summarised below. Lemma 6. Let k ≥ 2. nestn,k belongs to IAk2 , has polynomial size in n and is β-reducible to f Exp(k−1,n) x.

Contextual Approximation and Higher-Order Procedures

175

Note that the nested applications of f in f Exp(k−1,n) x are akin to generating a stack of height Exp(k − 1, n). We shall exploit this in our encodings. Note that, by substituting λccom .c; c for f in f Exp(k−1,n) x, we obtain a term that iterates x as many as Exp(k, n) times, i.e. Exp(k − 1, n)-fold nesting is used to simulate Exp(k, n)-fold iteration. Simulating Turing Machines. Let w be an input word. Let n = |w|, l = Exp(k − 1, n) and N = Exp(k, n). We shall consider a deterministic Turing machine T running in SPACE (l) and TIME (N ) and simulate T ’s behaviour on w. This suffices to establish SPACE (l)-hardness. We start off with the description of an encoding scheme for configurations of T . We shall represent them as strings of length l over an alphabet ΣT , to be specified later. We shall write ConfigT for the subset of (ΣT )l corresponding to configurations. The encoding of the initial configuration will be denoted by cinit and we shall write AcceptT for the set of representations of accepting configurations. Given c ∈ ConfigT , we write next(c) for the representation of the successor of c according to T ’s transition function. Let us introduce a number of auxiliary languages that will play an important role in the simulation. We write cR for the reverse of c. Definition 13. Let ΣT# = ΣT + {#}. We define the languages L0 , L1 ⊆ (ΣT )∗ and L2 , L3 , L4 ⊆ (ΣT# )∗ as follows. L0 = {cinit } L1 = AcceptT L2 = {cR # next(c) | c ∈ ConfigT } R L3 = {c # next(c) | c ∈ ConfigT } L4 = {c # dR | c ∈ ConfigT , d = next(c)} Lemma 7. There exists a representation scheme for configurations of T such that ΣT is polynomial in the size of T, w and the following properties hold. 1. There exist deterministic finite-state automata A0 , A1 , constructible from T, w in polynomial time, such that L(A0 ) ∩ (ΣT )l = L0 and L(A1 ) ∩ (ΣT )l = L1 . 2. For any i = 2, 3, 4, there exists a deterministic pushdown automaton Ai , constructible from T, w in polynomial time, such that L(Ai ) ∩ ((ΣT )l #(ΣT )l ) = Li . Moreover, transitions of the automata are given by three transition functions δpush : Qpush × ΣT → Qpush × Υ , δnoop : Qpush × {#} → Qpop and δpop : Qpop × ΣT × Υ → Qpop , the initial state belongs to Qpush and the automaton accepts by final state. I.e., the automata will process elements of (ΣT )l #(ΣT )l by performing push-moves first, then a noop move for # and, finally, pop-moves. Remark 3. Note that in the above lemma we had to use intersection with (ΣT )l (resp. (ΣT )l #(ΣT )l ) to state the correctness conditions with respect to ConfigT , because the automata will not be able to count up to l. However, in our argument, we are going to use the nesting power of IAk1 to run their transition functions for suitably many steps (l and 2l + 1 respectively). The significance of the languages L0 , L1 , L2 , L3 , L4 stems from the fact that they are building blocks of two other languages, L5 and L6 , which are closely related to the acceptance of w by T .

176

R. Lazi´c and A.S. Murawski

Lemma 8. Consider the languages L5 , L6 ⊆ (ΣT# )∗ defined by L5 = R R | cj ∈ ConfigT , f ∈ AcceptT , ∀i next(ci ) = {cinit # cR 1 # d1 # · · · cN # dN # f R di } and L6 = {c1 # d1 # · · · cN # dR N | cj ∈ ConfigT , ∃i next(ci ) = di }. Then T accepts w if and only if L5 ⊆ L6 . Proof. Note that if T accepts w then the sequence of (representations of the) configurations belonging to the accepting run, in which every other representation is reversed, gives rise to a word that belongs to L5 but not to L6 . R R ∈ L5 does not Conversely, if a word cinit # cR 1 # d 1 # · · · cN # d N # f belong to L6 then c1 = next(cinit ), ci+1 = next(di ) (i = 1, · · · , N − 1) and   f = next(dN ). Thus, the word actually represents an accepting run on w. Our hardness argument consists in translating the above lemma inside IAk1 . To that end, we shall show how to capture L2 , L3 , L4 and, ultimately, L5 and L6 , using IAk1 terms. We shall work under the assumption that ΣT# = {0, · · · , max }. Note, though, that the results can be adapted to any max > 0 by encoding ΣT# as sequences of exp-values. Similarly, using multiple exp-valued variables, IA-terms can store values that are bigger than max . We shall take advantage of such storage implicitly (e.g. for state values or stack elements), but the number of extra variables needed for this purpose will remain polynomial. Definition 14. We shall say that an IA-term z : exp  M : com captures L ⊆ (ΣT# )∗ if comp ([[z  M ]]) = {run q z (a1 )z q z (a2 )z · · · q z (ak )z done | a1 a2 · · · ak ∈ L}. Example 3. The term z : exp  M# : com, where M# ≡ if z = # then skip else Ω, captures {#}. In our constructions we often write [condition] to stand in for the assertion if (condition) then skip else Ω. Lemma 9. There exist IAk1 -terms z : exp  M0 , M1 : com, constructible from T, w in polynomial time, capturing L0 , L1 respectively. Lemma 10. There exists an IAk1 -term z : exp  M2 : com, constructible from T, w in polynomial time, which captures L2 . Thanks to the last two lemmas we are now ready to capture L5 . Lemma 11. There exists an IAk1 -term z : exp  M5 : com, constructible in polynomial time from T, w, which captures L5 . Proof. Note that a word from L5 contains N = Exp(k, n) segments from L2 . To account for that, it suffices to use N copies of M# ; M2 . However, for a polynomial-time reduction, we need to do that succinctly. Recall that nestn,k gives us l-fold nesting of functions, where l = Exp(k − 1, n). Consequently, N fold iteration can be achieved by l-fold nesting of λccom .c; c. Thus, we can take M5 ≡ M0 ; nestn,k [λccom .c; c/f, (M# ; M2 )/x]; M# ; M1 . To complete the hardness argument (by restating Lemma 8 using IAk1 terms), we also need to capture L6 . Because of the existential clause in its definition we need to use a slightly different capture scheme.

Contextual Approximation and Higher-Order Procedures

177

Lemma 12. There exists an IAk1 -term z : exp, FLAG : var  M6 : com, constructible in polynomial time from T, w, such that comp ([[z, FLAG  M6 ]]) = {run q z (a1 )z q z (a2 )z · · · q z (ak )z done | a1 a2 · · · ak ∈ L3 } ∪ {run q z (a1 )z q z FLAG (a2 )z · · · q z (ak )z write(1 ) ok FLAG done | a1 a2 · · · ak ∈ L4 }. Lemma 13. There exists an IAk1 -term z : exp  M6 : com, constructible in polynomial time from T, w, which captures L6 . Proof. It suffices to run M6 for N + 1 steps and check whether the flag was set: M6 ≡ new FLAG in (FLAG := 0; nestn,k [λccom .c; c/f, (M6 ; M# )/x]; M6 ; [!FLAG = 1])

Theorem 3. Contextual approximation between IAk1 terms is (k − 1)EXPSPACE-hard. Proof. By Lemmas 8, 11 and 13, for any Turing machine T running in SPACE (Exp(k − 1, n)) and TIME (Exp(k, n)) and an input word w, there exist IAk1 -terms x : exp  M5 , M6 , constructible from T, w in polynomial time, such that T accepts w if and only if M5 does not approximate M6 . This implies (k − 1)-EXPSPACE-hardness.  

6

Conclusion

We have shown that contextual approximation in IAk1 , IAk2 , IAk3 is (k − 1)EXPSPACE-complete. The algorithm that leads to these optimal bounds reduces terms to IA23 (with possibly (k − 2)-fold exponential blow-up) and we use a dedicated EXPSPACE procedure for IA23 exploiting game semantics and decision procedures for a special kind of pushdown automata. In particular, the results show that untamed β-reduction would yield suboptimal bounds, but selective β-reduction of redexes up to level 3 does not jeopardise complexity. The bounds above apply to open higher-order terms, i.e. IAi (i > 0) terms, for which the problem of contextual approximation is difficult to attack due to universal quantification over contexts. Our work also implies bounds for contextual approximation of IAk0 terms, i.e. closed terms of base type. Conceptually, this case is much easier, because it boils down to testing termination. In this case our techniques can be employed to obtain better upper bounds for IAk0 than those for IAk1 ((k − 1)-EXPSPACE). For a start, like for IAk1 , we can reduce IAk0 terms (at (k − 2)-fold exponential cost) to IA20 . Then termination in IA20 can be checked in exponential time by constructing pushdown automata via Lemma 2 and testing them for emptiness (rather than inclusion). Since emptiness testing of pushdown automata can be performed in polynomial time and the automata construction in Lemma 2 costs a single exponential, this yields an EXPTIME upper bound for termination in IA20 . Consequently, termination in IAk0 (k ≥ 2) can be placed in (k − 1)-EXPTIME, though it is not clear to us whether this bound is optimal. For completeness, let us just mention that termination in IA00 and IA10 is PSPACE-complete due to

178

R. Lazi´c and A.S. Murawski

presence of variables and looping (membership follows from the corresponding upper bounds for contextual equivalence). Another avenue for future work is IAk1 , IAk2 , IAk3 contextual equivalence. Of course, our upper bounds for approximation also apply to contextual equivalence, which amounts to two approximation checks. However, one might expect better bounds in this case given that our hardness argument leans heavily on testing inclusion. Finally, one should investigate how our results can be adapted to the callby-value setting. An educated guess would be that, in the analogous fragment of ML, the reduction of redexes up to order 3 (rather than 2) should be suppressed in order to obtain accurate complexity estimates.

References 1. Abramsky, S., McCusker, G.: Linearity, sharing, state: a fully abstract game semantics for Idealized Algol with active expressions. In: O’Hearn, P.W., Tennent, R.D. (eds.) Algol-Like Languages, pp. 297–329. Birkha¨ user, Boston (1997) 2. Abramsky, S., McCusker, G.: Game semantics. In: Schwichtenberg, H., Berger, U. (eds.) Logic and Computation, Proceedings of the 1997 Marktoberdorf Summer School. Springer-Verlag (1998) 3. Ahmed, A., Dreyer, D., Rossberg, A.: State-dependent representation independence.In: Proceedings of POPL, pp. 340–353. ACM (2009) 4. Alur, R., Madhusudan, P.: Visibly pushdown languages. In: Proceedings of STOC 2004, pp. 202–211 (2004) 5. Colvin, R., Hayes, I.J., Strooper, P.A.: Calculating modules in contextual logic program refinement. Theory Pract. Logic Program. 8(01), 1–31 (2008) 6. Ghica, D.R., McCusker, G.: Reasoning about idealized ALGOL using regular languages. In: Welzl, E., Montanari, U., Rolim, J.D.P. (eds.) ICALP 2000. LNCS, vol. 1853, p. 103. Springer, Heidelberg (2000) 7. Harmer, R.: Games and full abstraction for non-deterministic languages. Ph.D. thesis, University of London (2000) 8. Hopkins, D., Murawski, A.S., Ong, C.-H.L.: A fragment of ML decidable by visibly pushdown automata. In: Aceto, L., Henzinger, M., Sgall, J. (eds.) ICALP 2011, Part II. LNCS, vol. 6756, pp. 149–161. Springer, Heidelberg (2011) 9. Murawski, A.S.: On program equivalence in languages with ground-type references. In: Proceedings of IEEE Symposium on Logic in Computer Science, pp. 108–117. Computer Society Press (2003) 10. Murawski, A.S.: Games for complexity of second-order call-by-name programs. Theor. Comput. Sci. 343(1/2), 207–236 (2005) 11. Murawski, A.S., Walukiewicz, I.: Third-order idealized algol with iteration is decidable. In: Sassone, V. (ed.) FOSSACS 2005. LNCS, vol. 3441, pp. 202–218. Springer, Heidelberg (2005) 12. Pitts, A.M.: Operational semantics and program equivalence. In: Barthe, G., Dybjer, P., Pinto, L., Saraiva, J. (eds.) APPSEM 2000. LNCS, vol. 2395, pp. 378–412. Springer, Heidelberg (2002)

Contextual Approximation and Higher-Order Procedures

179

13. Reynolds, J.C.: The essence of Algol. In: de Bakker, J.W., van Vliet, J.C. (eds.) Algorithmic Languages, pp. 345–372. North Holland, Amsterdam (1978) 14. Turon, A., Dreyer, D., Birkedal, L.: Unifying refinement and hoare-style reasoning in a logic for higher-order concurrency. In: Proceedings of ICFP 2013, pp. 377–390 (2013)

A Theory of Slicing for Probabilistic Control Flow Graphs Torben Amtoft1(B) and Anindya Banerjee2 1

Kansas State University, Manhattan, KS, USA [email protected] 2 IMDEA Software Institute, Madrid, Spain [email protected]

Abstract. We present a theory for slicing probabilistic imperative programs—containing random assignment and “observe” statements— represented as control flow graphs whose nodes transform probability distributions. We show that such a representation allows direct adaptation of standard machinery such as data and control dependence, postdominators, relevant variables, etc. to the probabilistic setting. We separate the specification of slicing from its implementation: first we develop syntactic conditions that a slice must satisfy; next we prove that any such slice is semantically correct; finally we give an algorithm to compute the least slice. A key feature of our syntactic conditions is that they involve two disjoint slices such that the variables of one slice are probabilistically independent of the variables of the other. This leads directly to a proof of correctness of probabilistic slicing.

1

Introduction

The task of program slicing [12,14] is to remove the parts of a program that are irrelevant in a given context. This paper addresses slicing of probabilistic imperative programs which, in addition to the usual control structures, contain “random assignment” and “observe” statements. The former assign random values from a given distribution to variables. The latter remove undesirable combinations of values, a feature which can be used to bias the variables according to real world observations. The excellent survey by Gordon et al. [6] depicts many applications of probabilistic programs. Program slicing of deterministic imperative programs is increasingly well understood [1, 3,5, 10,11]. A basic notion is that if the slice contains a program point which depends on some other program points then these also should be included in the slice; here “depends” typically encompasses data dependence and control dependence. However, Hur et al. [7] recently demonstrated that in the presence of random assignments and observations, standard notions of data and A. Banerjee—Research supported by the US National Science Foundation (NSF). Any opinion, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of NSF. c Springer-Verlag Berlin Heidelberg 2016  B. Jacobs and C. L¨ oding (Eds.): FOSSACS 2016, LNCS 9634, pp. 180–196, 2016. DOI: 10.1007/978-3-662-49630-5 11

A Theory of Slicing for Probabilistic Control Flow Graphs

181

control dependence no longer suffice for semantically correct (backward) slicing. They develop a denotational framework in which they prove correct an algorithm for program slicing. In contrast, this paper shows how classical notions of dependence can be extended to give a semantic foundation for the (backward) slicing of probabilistic programs. The paper’s key contributions are: – A formulation of probabilistic slicing in terms of probabilistic control flow graphs (Sect. 3) that allows direct adaptation of standard machinery such as data and control dependence, postdominators, relevant variables, etc. to the probabilistic setting. We also provide a novel operational semantics of probabilistic control flow graphs (Sect. 4): written (v, D) ⇒ (v  , D ), the semantics states that as “control” moves from node v to node v  in the CFG, the probability distribution D transforms to distribution D . – Syntactic conditions for correctness (Sect. 5) that in a non-trivial way extend classical work on program slicing [5], and whose key feature is that they involve two disjoint slices; in order for the first to be a correct final result of slicing, the other must contain any “observe” nodes sliced away and all nodes on which they depend. We show that the variables of one slice are probabilistically independent of the variables of the other, and this leads directly to the correctness of probabilistic slicing (Sect. 6). – An algorithm, with running time at most cubic in the size of the program, that computes the best possible slice (Sect. 7) in that it is contained in any other (syntactic) slice of the program. Our approach separates the specification of slicing from algorithms to compute the best possible slice. The former is concerned with defining what is a correct syntactic slice, such that the behavior of the sliced program is equivalent to that of the original. The latter is concerned with how to compute the best possible syntactic slice; this slice is automatically a semantically correct slice—no separate proof is necessary. A program’s behavior is its final probability distribution; we demand equality modulo a constant factor so as to allow the removal of “observe” statements that do not introduce any bias in the final distribution. This will be the case if the variables tested by “observe” statements are independent, in the sense of probability theory, of the variables relevant for the final value. Full proofs of all results appear in the accompanying technical report [2].

2

Motivating Examples

Probabilistic Programs. Whereas in deterministic languages, a variable has only one value at a given time, we consider a language where a variable may have many different values at a given time, each with a certain probability. (Determinism is a special case where one value has probability one, and all others have probability zero.) We assume, to keep our development simple, that each possible value is an integer. A more general development, somewhat orthogonal to the aims of this paper, would allow real numbers and would require us to employ measure

182

T. Amtoft and A. Banerjee

theory (as explained in [9]); we conjecture that much will extend naturally (with summations becoming integrals). Similarly to [6], probabilities are introduced by the construct x := Random(ψ) which assigns to variable x a value with probability given by the random distribution ψ which in our setting is a mapping from Z (the set of integers) to ψ(z) = 1. A program phrase transforms a distribution into [0, 1] such that z∈Z

another distribution, where a distribution assigns a probability to each possible store. This was first formalized by Kozen [8] in a denotational setting. As also in [6], we shall use the construct Observe(B) to “filter out” values which do not satisfy the boolean expression B. That is, the resulting distribution assigns zero probability to all stores not satisfying B, while stores satisfying B keep their probability. The Examples. Slicing amounts to picking a set Q of “program points” (satisfying certain conditions as we shall soon discuss) and then removing nodes not in Q (as we shall formalize in Sect. 4.6). The examples all use a random distribution ψ4 over {0, 1, 2, 3} where ψ4 (0) = ψ4 (1) = ψ4 (2) = ψ4 (3) = 14 whereas / {0, 1, 2, 3}. The examples all consider whether it is correct ψ4 (i) = 0 for i ∈ to let Q contain exactly x := Random(ψ4 ) and Return(x), and thus slice into a program Px with straightforward semantics: after execution, the probability of each possible store is given by the distribution Δ defined as Δ ({x → i}) = 14 if i ∈ {0, 1, 2, 3}; otherwise Δ ({x → i}) = 0. def

Example 1. Consider the program P1 = 1: x := Random(ψ4 ); 2: y := Random(ψ4 ); 3: Observe(y ≥ 2); 4: Return(x). The distribution produced by 1 1 1 to each possible store the first two assignments will assign probability · = 4 4 16 {x → i, y → j} with i, j ∈ {0, 1, 2, 3}. In the final distribution D1 , a store {x → i, y → j} with j < 2 is impossible, and for each i ∈ {0, 1, 2, 3} there are thus only two possible stores that associate x with i: the store {x → i, y → 2}, and the store {x → i, y → 3}. Restricting to the variable x that is ultimately returned, 3  1 1 1 + = D1 ({x → i, y → j}) = D1 ({x → i}) = 16 16 8 j=2 if i ∈ {0, 1, 2, 3} (otherwise, D1 ({x → i}) = 0). We see that the probabilities in D1 do not add up to 1 which reflects that the purpose of an Observe statement is to cause undesired parts of the “local” distribution to “disappear” (which may give certain branches more relative weight than other branches). We also see that D1 equals Δ except for a constant factor: D1 = 0.5 · Δ . That is, Δ gives the same relative distribution over the values of x as D1 does. (An alternative way of phrasing this is that “normalizing” the “global” distribution, as done in [6], gives the same result for Px as for P1 .) We shall therefore say that Px is a correct slice of P1 . Thus the Observe statement is irrelevant to the final relative distribution of x. This is because y and x are independent in D1 , as formalized in Definition 3.

A Theory of Slicing for Probabilistic Control Flow Graphs

183

def

Example 2. Consider the program P2 = 1: x := Random(ψ4 ); 2: y := Random(ψ4 ); 3: Observe(x + y ≥ 5); 4: Return(x). Here the final distribution D2 allows only 3 stores: {x → 2, y → 3}), {x → 3, y → 2}) and {x → 3, y → 3}), 1 1 1 , and hence D2 ({x → 2}) = and D2 ({x → 3}) = . all with probability 16 16 8 Thus the program is biased towards high values of x; in particular we cannot write D2 in the form cΔ . Hence it is incorrect to slice P2 into Px . In this example, x and y are not independent in D2 ; this is as expected since the Observe statement in P2 depends on something (the assignment to x) on which the returned variable x also depends. def

Example 3. Consider the program P3 = Random(ψ4 ); Observe(z ≥ 3)); Return(x). values of x, with the final distribution D3 i ∈ {0, 1} and D3 ({x → i}) = D3 ({x → i, z it is incorrect to slice P3 into Px .

x := Random(ψ4 ); (if x ≥ 2 z := P3 is biased towards returning low given by D3 ({x → i}) = 14 when 1 → 3}) = 16 when i ∈ {2, 3}. Hence

The Observe statement cannot be removed: it is control dependent on the assignment to x, on which the returned x also depends. The discussion so far suggests the following tentative correctness condition for the set Q picked by slicing: – Q is “closed under dependency”, i.e., if a program point in Q depends on another program point then that program point also belongs to Q; – Q is part of a “slicing pair”: any Observe statement that is sliced away belongs to a set Q0 that is also closed under dependency and is disjoint from Q. The above condition will be made precise in Definition 9 (Sect. 5) which contains a further requirement, necessary since an Observe statement may be encoded as a potentially non-terminating loop, as the next example illustrates. def

Example 4. Consider the program P4 = x := Random(ψ4 ); y := Random(ψ4 ); (if x ≥ 2 (while y ≤ 5 do y := E)); Return(x) where E is an arithmetic expression. If E is “y + 1” then the loop terminates and y’s final value is 6. In the resulting distribution D , for i ∈ {0, 1, 2, 3} we have D ({x → i}) = D ({x → i, y → 6}) = 14 = Δ ({x → i}). Thus it is correct to slice P4 into Px . But if E is “y − 1” then the program will not terminate when x ≥ 2 (and hence the conditional encodes Observe(x < 2)). Thus the resulting distribution D4 is given by D4 ({x → i}) = 14 when i ∈ {0, 1} and D4 ({x → i}) = 0 when i∈ / {0, 1}. Thus it is incorrect to slice P4 into Px . Indeed, Definition 9 rules out such a slicing.

3

Control Flow Graphs

This section precisely defines the kind of CFGs we consider, as well as some key concepts that are mostly standard (see, e.g., [3, 10]). However, we also introduce a notion (Definition 1) specific to our approach.

184

T. Amtoft and A. Banerjee

Figure 1 depicts, with the nodes numbered, the CFGs corresponding to the programs P3 and P4 from Examples 3 and 4. We see that a node can be labeled with an assignment x := E (x a program variable and E an arithmetic expression), with a random assignment x := Random(ψ) (we shall assume that the probability distribution ψ contains no program variables though it would be straightforward to allow it as in [7]), with Observe(B) (B is a boolean expression), or (though not part of these examples) with Skip. Also, there are branching nodes with two outgoing edges. Finally, there is a unique End node Return(x) to which there must be a path from all other nodes but which has no outgoing edges, and a special node Start (which may have any label and is numbered 1 in the examples) from which there is a path to all other nodes. We let Def (v) be the variable occurring on the left hand side if v is a (random) assignment, and let Use(v) be the variables occurring in the right hand side of an assignment, in a boolean expression used in an Observe node or in a branching node, or as the sole variable in a End node. We demand that all variables be defined before they are used.

1

1 1: x := Random(ψ4 )

2 3

T

F

3: z := Random(ψ4 )

5

2: y := Random(ψ4 ) T

4: Observe(z ≥ 3)

3: x ≥ 2

3 F

4

5: Return(x)

4

1: x := Random(ψ4 )

2

2: x ≥ 2

T 5

F

6

4: y ≤ 5 5: y := E 6: Return(x)

Fig. 1. The CFGs for P3 (left) and P4 (right) from Examples 3 and 4.

We say that v1 postdominates v if v1 occurs on all paths from v to End; if also v1 = v, v1 is a proper postdominator of v. And we say that v1 is the first proper postdominator of v if whenever v2 is another proper postdominator of v then all paths from v to v2 contain v1 . It is easily shown that for any v with v = End, there is a unique first proper postdominator of v, called 1P P D(v). In Fig. 1(right), 1P P D(1) = 2, while also nodes 3 and 6 are proper postdominators. dd

We say that v2 is data dependent on v1 , written v1 → v2 , if there exists x ∈ Use(v2 ) ∩ Def (v1 ), and there exists a non-trivial path π from v1 to v2 such dd

that x ∈ / Def (v) for all nodes v that are interior in π. In Fig. 1(left), 1 → 2. A set dd

of nodes Q is closed under data dependence if whenever v2 ∈ Q and v1 → v2 then also v1 ∈ Q. We say that x is (Q-)relevant in v, written x ∈ rv Q (v), if there exists v  ∈ Q such that x ∈ Use(v  ), and a path π from v to v  such that x ∈ / Def (v1 ) for all v1 ∈ π \ {v  }. In Fig. 1(left), rv {4,5} (4) = {x, z} but rv {4,5} (3) = {x}. Next, a concept we have discovered useful for the subsequent development:

A Theory of Slicing for Probabilistic Control Flow Graphs

185

Definition 1. With v  a postdominator of v, and Q a set of nodes, we say that v stays outside Q until v  iff whenever π is a path from v to v  where v  occurs only at the end, π will contain no node in Q except possibly v  . In Fig. 1(right), node 4 stays outside {1, 6} until 6 but does not stay outside {1, 5, 6} until 6. It turns out that if v stays outside Q until v  and Q is closed under data dependence then v has the same Q-relevant variables as v  . Moreover, if Q satisfies certain additional properties, the distribution at v  (of the relevant variables) will equal the distribution at v.

4

Semantics

In this section we shall define the meaning of the CFGs introduced in the previous section, in terms of an operational semantics that manipulates distributions which assign probabilities to stores (Sect. 4.1). Section 4.2 defines what it means for sets of variables to be independent wrt. a given distribution. To prepare for the full semantics (Sect. 4.5) we define a one-step reduction (Sect. 4.3) from which we construct a reduction which allows multiple steps but only a bounded number of iterations (Sect. 4.4). The semantics also applies to sliced programs and hence (Sect. 4.6) provides the meaning of slicing. 4.1

Stores and Distributions

Let U be the universe of variables. A store s is a partial mapping from U to Z. We write s[x → z] for the store s that is like s except s (x) = z, and write dom(s) for the domain of s. We write S(R) for the set of stores with domain R, and also write F for S(U). If s1 ∈ S(R1 ) and s2 ∈ S(R2 ) with R1 ∩ R2 = ∅, we may define s1 ⊕ s2 with domain R1 ∪ R2 the natural way. With R a subset of U, R we say that s1 agrees with s2 on R, written s1 = s2 , iff R ⊆ dom(s1 ) ∩ dom(s2 ) and for all x ∈ R, s1 (x) = s2 (x). We assume that there is a function [[ ]] such that [[E]]s is the integer result of evaluating E in store s and [[B]]s is the boolean result of evaluating B in store s (the free variables of E, B must be in dom(s)). A distribution D (we shall the letter Δ) is a mapping  later also use  from F to non-negative reals with D < ∞ where D is a shorthand for s∈F D(s). Thanks to our assumption that valuesare integers, and since U can be assumed finite, F isa countable set and thus D is well-defined even without measure theory. If D ≤ 1, implying D(s) ≤ 1 for all s, we say that D is a probability distribution. We define D1 + D2 by stipulating (D1 + D2 )(s) = D1 (s) + D2 (s), and for c ≥ 0 we define cD by stipulating (cD)(s) = cD(s). We write D1 ≤ D2 iff D1 (s) ≤ D2 (s) for all s, and say that D = 0 iff D(s) =0 for all s. We assume DI = 1 (DI may be there is a designated initial distribution, DI , such that arbitrary as all variables must be defined before they are used). As suggested by the calculation in Example 1, we have  D(s0 ). Definition 2. For partial store s with domain R, let D(s) = R s0 ∈F | s=s0

186

T. Amtoft and A. Banerjee

Observe that D(∅) =



R

D. Say that D1 agrees with D2 on R, written D1 = D2 , R

R

if D1 (s) = D2 (s) for all s ∈ S(R). If D1 = D2 and R ⊆ R then D1 = D2 . 4.2

Probabilistic Independence

Some variables of a distribution D may be independent of others. Formally: Definition 3 (independence). Let R1 and R2 be disjoint sets of variables. We say that R1 and R2 are  independent in D iff for all s1 ∈ S(R1 ) and s2 ∈ S(R2 ), we have D(s1 ⊕ s2 ) D = D(s1 )D(s2 ).  To motivate the definition, first observe that if D = 1 it amounts  to the wellknown definition of probabilistic independence; next observe that if D > 0, it is equivalent to the well-known definition for “normalized” probabilities: D(s1 ) D(s2 ) D(s1 ⊕ s2 )  =  ·  D D D Trivially, R1 and R2 are independent in D if D = 0 or R1 = ∅ or R2 = ∅. Example 5. In Example 1, {x} and {y} are independent in D1 . This is since 1 so that for i ∈ {0, 1, 2, 3} and j ∈ {2, 3} we have D1 ({x → i, y → j}) = 16  1 1 1 D1 = ; we thus have the desired D1 ({x → i}) = , D1 ({y → j}) = , and 8 2  4 1 = D1 ({x → i}) · D1 ({y → j}). equality D1 ({x → i, y → j}) D1 = 32 And the equality holds trivially if i ∈ / {0, 1, 2, 3} or j ∈ / {2, 3} since then D1 ({x → i, y → j}) = 0 and either D1 ({x → i}) = 0 or D1 ({y → j}) = 0. Example 6. In Example 2,  {x} and3 {y} are not independent in D2 . This4 is while D2 ({x → 3})D2 ({y → 3}) = 256 . since D2 ({x → 3, y → 3}) D2 = 256 4.3

One-Step Reduction

If v has label Branch(B), with v1 (v2 ) the successor taken when B is true (false), we define T

– (v, D) → (v1 , D1 ) where D1 (s) equals D(s) when [[B]]s but is 0 otherwise; F – (v, D) → (v2 , D2 ) where D2 (s) is 0 when [[B]]s but equals D(s) otherwise. Thus D = D1 + D2 . Given v such that v has exactly one successor v  , the one step reduction (v, D) → (v  , D ) is given by defining D (s ) as follows: x := E

Skip D(s )

 s∈F |

x := Random(ψ)

s =s[x→[[E]]s] D(s)

If (v, D) → (v  , D ) then



D ≤

 s∈F | s



U \{x}

= s

ψ(s (x))D(s)



Observe(B) D(s ) if [[B]]s 0 otherwise

D with equality if v is not an Observe node.

A Theory of Slicing for Probabilistic Control Flow Graphs

4.4

187

Multi-step Reduction and Loops

The key semantic relation is of the form (v, D) ⇒ (v  , D ) where v  postdominates v, saying that distribution D transforms to distribution D as “control” moves from v to v  along paths that may contain multiple branches and even loops but which do not contain v  until the end. To define that relation, we need an k auxiliary relation of the form (v, D) ⇒ (v  , D ) where k is a non-negative integer that bounds the number of times control is allowed to move “away” from the End node; if k = 1 then we only take into account paths that for each step get closer to the End node, but if k = 2 we also allow paths with one cycle, etc. (k = 0 corresponds to “⊥” in denotational semantics.) Our goal is to let D be a function of v, D, v  and k, and we do so by a definition that is inductive first in k, and next on the length of the longest acyclic path from v to v  (proving properties of the relation will involve a case analysis on the various clauses). k

Definition 4. (⇒) Given v and v  where v  postdominates v, and given k and k

D, (v, D) ⇒ (v  , D ) holds when D is defined as follows: 1. if k = 0 then D = 0; 2. otherwise, if v  = v then D = D; 3. otherwise, if with v  = 1P P D(v) we have v  = v  , we recursively first find k

k

D with (v, D) ⇒ (v  , D ) and next find D with (v  , D ) ⇒ (v  , D ); 4. otherwise, if v has exactly one successor, which must be v  , we let D be such that (v, D) → (v  , D ); T F 5. otherwise, if v has two successors with (v, D) → (v1 , D1 ) and (v, D) → (v2 , D2 ) (thus v  postdominates v1 and v2 ), we recursively find D1 and D2 k

k

such that (v1 , D1 ) ⇒1 (v  , D1 ) and (v2 , D2 ) ⇒2 (v  , D2 ), and let D = D1 + D2 . Here ki (i = 1, 2) is given as follows: if the longest acyclic path from vi to v  is shorter than the length of the longest acyclic path from v to v  then ki = k, otherwise ki = k − 1. Example 7. Consider the CFG for P4 (Fig. 1(right)) with E chosen as “y + 1” k

and with Dk such that (1, DI ) ⇒ (6, Dk ). If i ∈ {0, 1} and j ∈ {0, 1, 2, 3} then for 1 1 and thus Dk ({x → i}) = . But if all k ≥ 1 we have Dk ({x → i, y → j}) = 16 4 i ∈ {2, 3} then we have Dk ({x → i, y → 6}) = 0 if k ≤ 3; D4 ({x → i, y → 6}) = 1 2 3 16 (for y initially 3); D5 ({x → i, y → 6}) = 16 ; D6 ({x → i, y → 6}) = 16 ; 4 and Dk ({x → i, y → 6}) = 16 if k ≥ 7. k

Note also that for k > 0, (4, D) ⇒ (4, D ) only holds if D = D, since the only path from 4 to 4 where 4 does not occur until the end is the empty path. Still, the k cycle between nodes 4 and 5 is taken into account. For if (4, D) ⇒ (6, D ) then D = D1 + D2 where D = D1 + D2 (D2 is D restricted to states where y > 5) k−1

k−1

and D1 is such that for some D1 , (5, D1 ) ⇒ (4, D1 ) and (4, D1 ) ⇒ (6, D1 ). D is a monotone function of D and of k:

188

T. Amtoft and A. Banerjee k

k

Lemma 1 (monotonicity). If (v, D1 ) ⇒1 (v  , D1 ) and (v, D2 ) ⇒2 (v  , D2 ) with    k k1 ≤ k2 and D1 ≤ D2 then D1 ≤ D2 . If (v, D) ⇒ (v  , D ) then D ≤ D.    D can fail due to Observe nodes (cf. Examples Equality between D and 1–3), or due to infinite loops (cf. Example 4) which cause k to be eventually zero. 4.5

Top-Level Semantics

Definition 5. (⇒) Given (v, D), and v  which postdominates v, (v, D) ⇒ (v  , D ) holds when D is defined as follows: with Dk (for k ≥ 0) the unique k

distribution such that (v, D) ⇒ (v  , Dk ), let D = limk→∞ Dk (the limit is taken pointwise). Observe by Lemma 1 that Dk1 ≤ Dk2 when k1 ≤ k2 ; for each s we thus have D0 (s) ≤ D1 (s) ≤ . . . ≤ Dk (s) ≤ Dk+1 (s) ≤ . . . and hence it is well-defined to let D (s) = lim Dk (s). Also at top-level, monotonicity holds: k→∞

Lemma 2. If (v, D) ⇒ (v  , D ) then



D ≤



D.

Example 8. Continuing Example 7, we see that the limit D is given as follows: 1 if i ∈ {0, 1}, j ∈ {0, 1, 2, 3}; D ({x → i, y → j}) = 14 D ({x → i, y → j}) = 16  if i ∈ {2, 3}, j = 6; and D ({x → i, y → j}) = 0 otherwise. We may define the meaning of a CFG with End node Return(x) as λv.D ({x → v}) where D is such that (Start, DI ) ⇒ (Return(x), D ). 4.6

Semantics of Slicing

A slice set is a set of nodes (which must satisfy certain conditions, cf. Definition 9). Slicing amounts to ignoring nodes not in the slice set: Definition 6. (=⇒) Given a slice set Q, we define the semantics of the CFG that results from slicing wrt. Q as follows: k k Let (v, Δ) =⇒ (v  , Δ ) be defined as (v, D) ⇒ (v  , D ) (Definition 4), except that whenever v ∈ / Q ∪ {End} then v is labeled Skip and the successor of v becomes 1P P D(v). And let (v, Δ) =⇒ (v  , Δ ) be defined by letting Δ = lim Δk where for each k ≥ 0, Δk is the unique distribution such that k→∞ k

(v, Δ) =⇒ (v  , Δk ).

5

Conditions for Slicing

With Q the slice set, we now develop conditions for Q that ensure semantic correctness. It is standard to require Q to be closed under data dependence, and additionally also under some kind of “control dependence”. In this section we first elaborate on the latter condition after which we study the extra conditions

A Theory of Slicing for Probabilistic Control Flow Graphs

189

needed in our probabilistic setting. Eventually, Definition 9 gives conditions that involve not only Q but also another slice set Q0 containing all Observe nodes to be sliced away. As stated in Proposition 1, these conditions are sufficient to establish probabilistic independence of Q and Q0 . This in turn is crucial for establishing the correctness of slicing, as stated in Theorem 1 (Sect. 6). Weak Slice Sets. Danicic et al. [5] show that various kinds of control dependence can all be elegantly expressed within a general framework whose core is the following notion: Definition 7 (next observable). With Q a set of nodes, v  is a next observable in Q of v iff v  ∈ Q ∪ {End}, and v  occurs on all paths from v to a node in Q ∪ {End}. A node v can have at most one next observable in Q. It thus makes sense to write v  = next Q (v) if v  is a next observable in Q of v. We say that Q provides next observables iff next Q (v) exists for all nodes v. If v  = next Q (v) then v  is a postdominator of v, and if v ∈ Q ∪ {End} then next Q (v) = v. In the CFG for P3 (Fig. 1), letting Q = {1, 3, 5}, node 5 is a next observable in Q of 4: all paths from 4 to a node in Q will contain 5. But no node is a next observable in Q of 2: node 3 is not since there is a path from 2 to 5 not containing 3, and node 5 is not since there is a path from 2 to 3 not containing 5. Therefore Q cannot be the slice set: node 1 can have only one successor in the sliced program but we have no reason to choose either of the nodes 3 and 5 over the other as that successor. This motivates the following definition: Definition 8 (weak slice set). We say that Q is a weak slice set iff it provides next observables, and is closed under data dependence. While the importance of “provides next observable” was recognized already in [1,11], Danicic et al. were the first to realize that it is the key property (together with data dependence) to ensure semantically correct slicing. They call the property “weakly committing” (thus our use of “weak”) and our definition differs slightly from theirs in that we always consider End an “observable”. It is easy to see that the empty set, as well as the set of all nodes, is a weak slice set. Moreover, if Q1 , Q2 are weak slice sets, so is Q1 ∪ Q2 . Adapting to the Probabilistic Setting. The key challenge in slicing probabilistic programs is, as already motivated through Examples 1–4, to handle Observe nodes. In Sect. 2 we hinted at some tentative conditions a slice set Q must satisfy; we can now phrase them more precisely: 1. Q must be a weak slice set that contains End, and 2. there exists another weak slice set Q0 such that (a) Q and Q0 are disjoint and (b) all Observe nodes belong to either Q or Q0 . For programs P1 , P2 , the control flow is linear and hence all nodes have a next observable (so a node set that is closed under data dependence is a weak slice set).

190

T. Amtoft and A. Banerjee

For P1 we may choose Q = {1, 4} and Q0 = {2, 3} as they are disjoint, and both closed under data dependence. Hence we may use {1, 4} as a slice set; from Definition 6 we see that the resulting slice is 1: x := Random(ψ4 ); 2: Skip; 3: Skip; 4: Return(x), which is obviously equivalent to Px as defined in Sect. 2. Next consider the program P2 where Q should contain 4 and hence (by data dependence) also contain 1. Now assume, in order to remove the Observe node (and produce Px ), that Q does not contain 3. Then Q0 must contain 3, and (as Q0 is closed under data dependence) also 1. But then Q and Q0 are not disjoint, which contradicts our requirements. Thus Q does contain 3, and hence also 2. That is, Q = {1, 2, 3, 4}. We see that the only possible slicing is the trivial one. Any slice for P3 will also be trivial. From 5 ∈ Q we infer (by data dependence) that 1 ∈ Q. Assume, to get a contradiction, that 4 ∈ / Q. Then 4 ∈ Q0 , and for node 2 to have a next observable in Q0 we must also have 2 ∈ Q0 which by data dependence implies 1 ∈ Q0 which as 1 ∈ Q contradicts Q and Q0 being disjoint. Thus 4 ∈ Q which implies 3 ∈ Q (by data dependence) and 2 ∈ Q (as otherwise 2 has no next observable in Q). For P4 , it is plausible that Q = {1, 6} and Q0 = ∅, since for all v = 1 we would then have 6 = next Q (v). From Definition 6 we see that after removing unreachable nodes, the resulting slice is: 1: x := Random(ψ4 ); 2: Skip; 3: Skip; 6: Return(x). Yet, in Example 4 we saw that this is in general not a correct slice of P4 . This reveals a problem with our tentative correctness conditions; they do not take into account that Observe nodes may be “encoded” as infinite loops. To repair that, we demand that just like all Observe nodes must belong to either Q or Q0 , also all cycles must touch either Q or Q0 . With this requirement, it is no longer possible that Q contains only nodes 1 and 6. For if so, then Q0 would have to contain node 4 or node 5 since these two nodes form a cycle. But then, in order for node 3 to have a next observable in Q0 , it must be the case that Q0 contains node 3, and hence (by data dependence) also node 1 which contradicts Q and Q0 being disjoint. Thus we have motivated the following definition. Definition 9 (slicing pair). Let Q, Q0 be sets of nodes. (Q, Q0 ) is a slicing pair iff (a) Q, Q0 are both weak slice sets with End ∈ Q; (b) Q, Q0 are disjoint; (c) all Observe nodes are in Q ∪ Q0 ; and (d) all cycles contain at least one element of Q ∪ Q0 . If (Q, Q0 ) is a slicing pair then rv Q (v) ∩ rv Q0 (v) = ∅ for all nodes v. For, if x ∈ rv Q (v) ∩ rv Q0 (v) then a definition of x, known to exist, would be in Q ∩ Q0 which is empty. Moreover, the Q-relevant variables are probabilistically independent (as defined in Definition 3) of the Q0 -relevant variables, as stated by the following preservation result which is one of the main contributions of this paper in that it gives a syntactic condition for probabilistic independence: Proposition 1 (Independence). Let (Q, Q0 ) be a slicing pair. Assume that k

(v, D) ⇒ (v  , D ) where v  postdominates v. If rv Q (v) and rv Q0 (v) are independent in D then rv Q (v  ) and rv Q0 (v  ) are independent in D .

A Theory of Slicing for Probabilistic Control Flow Graphs

6

191

Slicing and Its Correctness

We can now precisely phrase the desired correctness result: Theorem 1. Given a CFG with End of the form Return(x), and let (Q, Q0 ) be a slicing pair. Let D and Δ be the unique distributions such that (Start, DI ) ⇒ (End, D ) and (Start, DI ) =⇒ (End, Δ ) where the latter denotes slicing wrt. Q (cf. Sect. 4.6). Then there exists a real number c with 0 ≤ c ≤ 1 such that for all v, D ({x → v}) = cΔ ({x → v}). This follows from a more general proposition (allowing an inductive proof) stated below, with v = Start so that (from the requirement that a variable must be defined before it is used) R and R0 are both empty (and thus independent), with v  = End = Return(x) so that R = {x}, and with Δ = D = DI . Proposition 2. Let (Q, Q0 ) be a slicing pair. Let v  postdominate v, with R = rv Q (v), R = rv Q (v  ), and R0 = rv Q0 (v). Assume that D is such that R and R

R0 are independent in D, and that Δ is such that D = Δ. Let D and Δ be the unique distributions such that (v, D) ⇒ (v  , D ) and (v, Δ) =⇒ (v  , Δ ). Then R

there exists a real number c with 0 ≤ c ≤ 1 such that D = cΔ . Moreover, c = 1 if v stays outside Q0 until v  (since then the conditions for a slicing pair guarantee that no observe nodes, or infinite loops, are sliced away). That is, for the relevant variables, the final distribution is the same in the sliced program as  in the original program, except for a constant factor c such that  D  = c Δ . To prove Proposition 2, we need a similar result that involves k and where the proof of sequential composition (case 3 in Definition 4) employs Proposition 1 to ensure that the independence property still holds after the first reduction, so that we can apply the induction hypothesis to the second reduction.

7

Computing the (Least) Slice

There always exists at least one slicing pair, with Q the set of all nodes and with Q0 the empty set; in that case, the sliced program is the same as the original. Our goal, however, is to find a slicing pair (Q, Q0 ) where Q is as small as possible while including the End node. This section describes an algorithm BSP for doing so. The running time of our algorithms is measured in terms of n, the number of nodes in the graph. Note that the number of edges is at most 2n and thus in O(n). Our algorithms use a boolean table DD∗ such that DD∗ (v, v  ) is true iff dd ∗

dd ∗

dd

v → v  where → is the reflexive and transitive closure of →. Given DD∗ , it is easy to ensure that sets are closed under data dependence, and we shall do that in an incremental way: there exists an algorithm DDclose which given a node set Q0 that is closed under data dependence, and a node set Q1 , returns the least set Q containing Q0 and Q1 that is closed under data dependence.

192

T. Amtoft and A. Banerjee

PN?(Q) F ← Q ∪ {End}; foreach v ∈ V if v ∈ F N [v] ← v else N [v] ← ⊥; C ← ∅; while F = ∅ ∧ C = ∅ F ← ∅; foreach edge from v ∈ /Q to v ∈ F if N (v) = ⊥ N (v) ← N (v ); F ← F ∪ {v} else if N (v) = N (v ) C ← C ∪ {v}; F ←F return C

LWS(Q0 ) Q ← DDclose (∅, Q0 ); C ← PN?(Q); while C = ∅ Q ← DDclose (Q, C); C ← PN?(Q) return Q BSP() W ← the set of essential nodes; foreach v ∈ W ∪ {End} Qv ← LWS({v}); Q ← ∅; F ← QEnd ; while F = ∅ Q ← Q ∪ F ; F ← ∅; foreach v ∈ W if Qv ∩ Q = ∅ W ← W \ {v}; F ← F ∪ Qv ; Q0 ← v∈W Qv ; return (Q, Q0 )

Fig. 2. Algorithms for: checking if a set provides next observables (PN?); finding the least weak slice set containing a given set (LWS); finding the best slicing pair (BSP).

Computing the Least Weak Slice. In Fig. 2(upper right) we define a function LWS which constructs the least weak slice set that contains a given set; it works by successively adding nodes to the set until it is closed under data dependence, and provides next observables. To check the latter, we employ a function PN? as defined in Fig. 2(left): given Q, it does a backward breadth-first search from Q ∪ {End} to find the first node (if any) from which two nodes in that set are reachable without going through Q; such a node must be included in any superset providing next observables. Lemma 3. Given Q, the function PN? runs in time O(n) and returns C such that (a) if C is empty then Q provides next observables, and (b) if C is non-empty then C ∩ Q = ∅ and all supersets of Q that provide next observables contain C. Lemma 4. Given Q0 , the function LWS returns Q such that Q is a weak slice set and Q0 ⊆ Q. Moreover, if Q is a weak slice set with Q0 ⊆ Q , then Q ⊆ Q . Finally, given DD∗ , LWS runs in time O(n2 ). Computing the Best Slicing Pair. We now develop an algorithm BSP which given a CFG returns a slicing pair (Q, Q0 ) with Q as small as possible. From Definition 9 we know that each Observe node has to be put either in Q or in Q0 , and also that at least one node from each cycle has to be put either in Q or in Q0 . Especially the latter requirement may suggest that our algorithm will have to explore a huge search space, but fortunately it is sufficient to consider only the node(s) with minimal height. Here node v’s height, denoted H(v), is the length of the shortest path(s) from v to End. This motivates

A Theory of Slicing for Probabilistic Control Flow Graphs

193

Definition 10. A node v is essential if either (a) v is an Observe node, or (b) v belongs to a cycle π where H(v) ≤ H(v1 ) for all v1 ∈ π. For disjoint weak slice sets (Q, Q0 ) with End ∈ Q to be a slicing pair, it is sufficient and necessary that each essential node be placed either in Q or in Q0 . Lemma 5 (Sufficient). Let Q and Q0 be disjoint weak slice sets with End ∈ Q, and assume that all essential nodes are in Q∪Q0 . Then (Q, Q0 ) is a slicing pair. Lemma 6 (Necessary). Let (Q, Q0 ) be a slicing pair. If v is essential then v ∈ Q ∪ Q0 . Fig. 2 (lower right) presents the algorithm BSP that computes the best slicing pair (Q, Q0 ). The idea is to built Q incrementally, initially containing only End, and then add v whenever v is essential but cannot be placed in Q0 as then Q and Q0 would overlap. That BSP produces the best slicing pair is captured by Proposition 3. The algorithm BSP returns (on a given CFG) Q and Q0 such that (Q, Q0 ) is a slicing pair, and if (Q , Q0 ) is a slicing pair then Q ⊆ Q . After analyzing the running times, we get: Theorem 2. For a given CFG, there is an algorithm that in time O(n3 ) computes a slicing pair (Q, Q0 ) where Q ⊆ Q for any other slicing pair (Q , Q0 ). Also the algorithm given in [5] for computing (their version of) weak slices runs in cubic time. We do not expect that there exists an algorithm with lower asymptotic complexity, since we need to compute data dependencies which is known to involve computing a transitive closure. Illustrating the Algorithms. First consider the program P1 from Example 1 where dd the non-trivial true entries of DD∗ are (1, 4) (since 1 → 4) and (2, 3), and where 3 is the only essential node. BSP thus computes LWS({4}) and LWS({3}). When running LWS on {4}, initially Q = {1, 4} which is also the final value of Q since PN?({1, 4}) returns ∅ (after a sequence of iterations where F is first {1, 4} and next {3} and next {2} and finally ∅). Thus Q4 = {1, 4} and similarly we get Q3 = {2, 3}. When the members of W = {3} are first examined in the BSP algorithm, we have Q = Q4 and thus Q3 ∩ Q = ∅. Hence the while loop terminates with Q = {1, 4} and subsequently we get Q0 = Q3 = {2, 3}. Next consider the CFG for P4 (Fig. 1) with E containing only y free. Here dd

dd

dd

dd

dd

dd

dd

→ is given as follows: 1 → 3, 1 → 6, 2 → 4, 2 → 5, 5 → 4, and 5 → 5. Since H(4) = 1 and H(5) = 2, node 4 is the only essential node. BSP thus has to compute LWS({6}) and LWS({4}): – LWS({6}): initially, Q = {1, 6} which is also the final value of Q since PN?({1, 6}) returns ∅ (after a sequence of iterations where F is first {1, 6} and next {3, 4} and next {2, 5} and finally ∅). Thus Q6 = {1, 6}.

194

T. Amtoft and A. Banerjee

– LWS({4}): initially, Q = {2, 4, 5}. In PN? we initially thus have F = {2, 4, 5, 6} which causes the first iteration of the while loop to put 3 in C so that {3} dd

is eventually returned. Since 1 → 3 holds, the next iteration of LWS will have Q = {1, 2, 3, 4, 5} on which PN? will return ∅. Thus Q4 = {1, 2, 3, 4, 5}. When the members of W = {4} are first examined in the BSP algorithm, we have Q = Q6 and thus Q4 ∩Q = {1} = ∅. Hence Q will equal Q6 ∪Q4 = {1, 2, 3, 4, 5, 6} and as W is now empty, the loop will terminate and we get Q0 = ∅.

8

Extensions and Future Work

Section 7 presented an algorithm for computing the least syntactic slice. But there may exist even smaller slices that are still semantically correct: recall Example 4 where the only correct syntactic slice is the program itself (as shown in Sect. 7) but where a much smaller slice may be semantically correct for certain instantiations of the generic “E”. Similarly, a node labeled Observe(B) can be safely discarded if B always evaluates to true. For example, the CFG with textual representation 1 : x := Random(ψ4 ); 2 : y := 7; 3 : if x ≥ 2 (4 : Observe(y = 7)); 5 : Return(x) is semantically equivalent to the CFG containing only nodes 1 and 5. But the former has no smaller syntactic slice, since if (Q, Q0 ) is a slicing pair with 5 ∈ Q (and thus 1 ∈ Q) then Q = {1, 2, 3, 4, 5} as we now show. If 4 ∈ Q0 then 3 ∈ Q0 (as Q0 provides next observables) and thus 1 ∈ Q0 which contradicts Q ∩ Q0 = ∅. Hence (as 4 must belong to Q ∪ Q0 ) 4 ∈ Q; but then 2 ∈ Q (by data dependence) and 3 ∈ Q (as Q provides next observables). Simple analyses like constant propagation may improve the precision of slicing even in a deterministic setting, but the probabilistic setting gives an extra opportunity: after an Observe(B) node, we know that B holds. As richly exploited in [7], a simple syntactic transformation often suffices to get the benefits of that information, as we illustrate on the program from [7, Fig. 4] whose CFG (in slightly modified form) is depicted in Fig. 3. In our setting, if (Q, Q0 ) with 18 ∈ Q is the best slicing pair, then Q will contain everything except nodes 12, 13, 14, as can be seen as follows: 16, 17 ∈ Q by data dependence; 15 ∈ Q as Q provides next observables; 6, 7, 8, 9 ∈ Q by data dependence; 3, 4, 5 ∈ Q as Q provides next observables; 1, 2 ∈ Q by data dependence; also 10 ∈ Q as otherwise 10 ∈ Q0 and thus also 9 ∈ Q0 which contradicts Q ∩ Q0 = ∅. Alternatively, suppose we insert a node 11 labeled g := 0 between nodes 10 and 12. This clearly preserves the semantics, but allows a much smaller slice: choose Q = {11, 15, 16, 17, 18} and Q0 = {1, 2, 3, 4, 5, 6, 7, 8, 9, 10}. This is much like what is arrived at (through a more complex process) in [7, Fig. 15]. Future work involves exploring a larger range of examples, and investigating useful techniques for computing slices that are smaller than the least syntactic slice yet semantically correct. (Of course it is in general undecidable to compute the least semantically correct slice.) We have recently augmented our theory such that we can ignore loops that are known (by some means) to always terminate.

A Theory of Slicing for Probabilistic Control Flow Graphs

195

18

1 d := R(..)

Ret(l)

2 i := R(..) 4

3

16

17

i=0

l := R(..)

l := R(..)

T

6

T

g := R(..)

F

5

F

d=0

T

T

7

8

g := R(..)

g := R(..)

F

F 15

d=0

g=0 9

g := R(..)

13

14

s := R(..)

s := R(..) T

F

10

12

Obs(g = 0)

i=0

Fig. 3. The program from [7, Fig. 4] (modified).

That is, for a slicing pair (Q, Q0 ), a cycle which cannot go on forever (or does it with probability zero) does not need to contain a node from Q ∪ Q0 .

9

Conclusion and Related Work

We have developed a theory for the slicing of probabilistic imperative programs. We have used and extended techniques from the literature [1, 3, 10, 11] on the slicing of deterministic imperative programs. These frameworks, some of which have been partly verified by mechanical proof assistants [4, 13], were recently coalesced by Danicic et al. [5] who provided solid semantic foundations to the slicing of a large class of deterministic programs. Our extension of that work is non-trivial in that we need to capture probabilistic independence as done in Proposition 1 which requires two slice sets rather than just one. We were directly inspired by Hur et al. [7] who point out the challenges involved in the slicing of probabilistic programs, and present an algorithm which constructs a semantically correct slice. The paper does not state whether it is in some sense the least possible slice; neither does it address the complexity of the algorithm. While Hur et al.’s approach differs from ours, for example it is based on a denotational semantics for a structured language (we expect the two semantics to coincide for CFGs of structured programs), it is not surprising that their correctness proof also has probabilistic independence (termed “decomposition”) as a key notion. Our theory separates specification and implementation which we believe provides for a cleaner approach. But as mentioned in Sect. 8, they incorporate powerful optimizations that we do not (yet) allow. Acknowledgements. Thanks to Gordon Stewart for comments on earlier drafts.

196

T. Amtoft and A. Banerjee

References 1. Amtoft, T.: Slicing for modern program structures: a theory for eliminating irrelevant loops. Inf. Process. Lett. 106(2), 45–51 (2008) 2. Amtoft, T., Banerjee, A.: A theory of slicing for probabilistic control flow graphs. Technical Report CIS TR 2015-1, Kansas State University, July 2015. http:// people.cis.ksu.edu/∼tamtoft/Papers/Amt+Ban:ProbSlice-2015/long.pdf 3. Ball, T., Horwitz, S.: Slicing programs with arbitrary control flow. In: Fritzson, P.A. (ed.) AADEBUG 1993. LNCS, vol. 749, pp. 206–222. Springer, Heidelberg (1993) 4. Blazy, S., Maroneze, A., Pichardie, D.: Verified validation of program slicing. In: Proceedings of the 2015 Conference on Certified Programs and Proofs, CPP 2015, pp. 109–117. ACM, New York (2015) ´ Laurence, 5. Danicic, S., Barraclough, R.W., Harman, M., Howroyd, J.D., Kiss, A., M.R.: A unifying theory of control dependence and its application to arbitrary program structures. Theor. Comput. Sci. 412(49), 6809–6842 (2011) 6. Gordon, A.D., Henzinger, T.A., Nori, A.V., Rajamani, S.K.: Probabilistic programming. In: Dwyer, M.B., Herbsleb, J. (eds.) ICSE, Future of Software Engineering track, FOSE 2014, pp. 167–181. ACM, New York (2014) 7. Hur, C.-K., Nori, A.V., Rajamani, S.K., Samuel, S.: Slicing probabilistic programs. In: Pingali, K. (ed.) Proceedings of the 35th ACM SIGPLAN Conference on Programming Language Design and Implementation, PLDI 2014, pp. 133–144. ACM, New York (2014) 8. Kozen, D.: Semantics of probabilistic programs. J. Comput. Syst. Sci. 22, 328–350 (1981) 9. Panangaden, P.: Labelled Markov Processes. Imperial College Press, London (2009) 10. Podgurski, A., Clarke, L.A.: A formal model of program dependences and its implications for software testing, debugging, and maintenance. IEEE Trans. Softw. Eng. 16(9), 965–979 (1990) 11. Ranganath, V.P., Amtoft, T., Banerjee, A., Hatcliff, J., Dwyer, M.B.: A new foundation for control dependence and slicing for modern program structures. ACM Trans. Program. Lang. Syst. (TOPLAS), 29(5), Aug 2007. (A special issue with extended versions of selected papers from the 14th European Symposium on Programming (ESOP’05)) 12. Tip, F.: A survey of program slicing techniques. J. Program. Lang. 3, 121–189 (1995) 13. Wasserrab, D.: From Formal Semantics to Verified Slicing. Ph.D. thesis, Karlsruher Institut f¨ ur Technologie (2010) 14. Weiser, M.: Program slicing. IEEE Trans. Softw. Eng. 10(4), 352–357 (1984)

Verification of Parameterized Communicating Automata via Split-Width Marie Fortin and Paul Gastin(B) LSV, ENS Cachan, CNRS, Universit´e Paris-Saclay, 94235 Cachan, France [email protected], [email protected]

Abstract. We study verification problems for distributed systems communicating via unbounded FIFO channels. The number of processes of the system as well as the communication topology are not fixed a priori. Systems are given by parameterized communicating automata (PCAs) which can be run on any communication topology of bounded degree, with arbitrarily many processes. Such systems are Turing powerful so we concentrate on under-approximate verification. We extend the notion of split-width to behaviors of PCAs. We show that emptiness, reachability and model-checking problems of PCAs are decidable when restricted to behaviors of bounded split-width. Reachability and emptiness are Exptime-complete, but only polynomial in the size of the PCA. We also describe several concrete classes of bounded split-width, for which we prove similar results. Keywords: Parameterized distributed systems Split-width · Message sequence charts

1

·

Model checking

·

Introduction

We study verification problems for parameterized communicating automata (PCAs), which model distributed systems consisting of arbitrarily many identical processes, distributed on some communication topology. Each process runs a copy of the same finite automaton, that can send and receive messages from other processes through FIFO channels. Though the system may contain unboundedly many processes, we assume that each process may only communicate with a bounded number of other processes. PCAs were introduced in [5] to study logical characterizations of parameterized systems. They extend communicating finite-state machines [8]. While the latter assume a fixed and known communication topology, a PCA can be run on any communication topology of bounded degree. Communicating finite-state machines are already Turing equivalent, and thus their verification is undecidable. The fact that PCAs can be run on arbitrarily large topologies induces other sources of undecidability. For instance, a Turing machine can be simulated on grid topologies by a PCA performing a bounded number of actions on each process. Thus, some restrictions on the topologies are necessary. Yet, even when c Springer-Verlag Berlin Heidelberg 2016  B. Jacobs and C. L¨ oding (Eds.): FOSSACS 2016, LNCS 9634, pp. 197–213, 2016. DOI: 10.1007/978-3-662-49630-5 12

198

M. Fortin and P. Gastin

fixing a simple class of topologies such as pipelines, and imposing rendez-vous synchronization, reachability of PCAs is undecidable [7]. In order to regain decidability, we focus on under-approximate verification. The idea is to restrict the verification problems to meaningful classes C of finite behaviors. Typically, we are interested in the following problem: given a PCA S, is there a topology T and a behavior in C over T on which S reaches some local/global state? Our aim is to study under-approximation classes C for which verification problems of PCAs becomes decidable. Even when we cannot cover all behaviors of a system with a decidable class, under-approximate verification is still very useful to detect bugs. Usually, the classes Ci are parameterized and cover more and more behaviors when the parameter i increases. The behaviors of PCAs are described by finite message sequence charts (MSC) [18], that is, graphs describing the communications between the different processes. Each node of the graph corresponds to an action performed by some process (sending or receiving a message), and the dependencies between the different actions are indicated by the edges of the graph. For model-checking, the specifications of our systems are typically given as monadic second order logic (MSO) or propositional dynamic logic (PDL) formulas over MSCs. It is known that for any MSO definable class of bounded degree graphs, having a decidable MSO theory is equivalent to having bounded tree-width [9]. This applies to classes of MSCs, and characterizes decidability of MSO modelchecking. However, showing a bound on tree-width is in general difficult. In the case of MSCs over a fixed architecture, an alternative notion called split-width has been introduced [1, 2]. Split-width is equivalent to tree-width on MSCs, but easier to use. It provides means to define uniform decision procedures, applying to many well-studied restrictions of communicating finite-state machines [10]. Following this approach, we generalize the definition of split-width, and we extend existing results over fixed architectures to the parameterized case. The idea of split-width is to decompose an MSC into atomic pieces, that is, pairs of matching send and receive events. This is done using two operations: splitting some edges between consecutive events of a same process, and dividing the resulting graph into disjoint components. Intuitively, an MSC has bounded splitwidth when this can be done while splitting a bounded number of edges at each step, on a bounded number of processes. We show that emptiness and reachability of PCAs restricted to MSCs of bounded split-width are decidable. These problems are Exptime-complete but only polynomial in the size of the PCA. Our decision procedures are based on an interpretation of MSCs of bounded split-width into binary trees, and reductions to tree automata verification problems. In the extended version [16] we also prove that model-checking restricted to bounded split-width is decidable. For MSO specifications, it has non-elementary complexity. However, the under-approximate model-checking is respectively Exptime-complete and 2-Exptime complete for CPDL (PDL with converse) and ICPDL (PDL with converse and intersection). In all cases, the problem is still polynomial in the size of the PCA.

Verification of Parameterized Communicating Automata via Split-Width

199

Further, we give several examples of concrete classes of MSCs with bounded split-width, including existentially bounded or context-bounded behaviors, for which we show similar decidability and complexity results. Our approach based on split-width is generic since it can be easily adapted to other under-approximation classes. Related Work. Various models of parameterized systems have been considered in the literature. In several cases, the assumptions of the model are sufficient to get decidability results without additional restrictions on the behaviors. Examples include token-passing systems [3, 15], models with a global store without locking [14], message-passing systems communicating synchronously via broadcast [12, 13], or rendez-vous [4]. An important distinction between the latter and PCAs is that PCAs use point-to-point communication: a process can distinguish between its neighbors, and specify the recipients of its messages. This makes the model more expressive, but also leads to undecidability. The emptiness and model-checking problems for PCAs have been considered in [6, 7], respectively. Both papers assume rendez-vous synchronization, and a fixed class of topologies: pipelines, rings, or trees. Several notions of contexts are introduced, and decision procedures are described for the corresponding classes of context-bounded behaviors. In the present paper, we also address context-bounded verification, but for unbounded FIFO channels and arbitrary topologies of bounded tree-width. In the general case of bounded split-width, we do not assume any restriction on the topology, but a bound on split-width already implies a bound on the tree-width of the communication topology. Communicating finite-state machines (over fixed topologies) have been more extensively studied, and several restrictions are known to bring back decidability. Our work generalizes some of them to the parameterized setting, namely, context bounds (introduced in [19] for multi-stack concurrent systems), existential bounds [17], and bounded split-width [2]. Split-width was first introduced for multi-pushdown systems [11], and then generalized to communicating multi-pushdown systems [2]. Outline. In Sect. 2, we define topologies, PCAs, and MSCs. In Sect. 3, we introduce split-width. In Sect. 4, we give several examples of classes of bounded split-width, and state our results for the reachability problems of those classes. In Sect. 5, we present in more details the decision procedures leading to these results. In Sect. 6, we briefly present how they can be extended to decide modelchecking problems, and discuss possible extensions of our model. Most proofs are omitted and can be found in the full version of the paper [16].

200

2

M. Fortin and P. Gastin

Parameterized Communicating Automata

We describe our formal model for communicating systems: we introduce topologies, MSCs, and parameterized communicating automata. Our definitions are taken from [5], except that we abstract away idle processes. Topologies. We model distributed systems consisting of an unbounded number of processes. Each process is equipped with a bounded number of interfaces, through which it can communicate with other processes. A topology is a graph, describing the connections between the different processes (each of which is represented by a node in the graph, see example on the right). Throughout the paper, we assume a fixed nonempty finite set N = {a, b, . . .} of interface names (or, simply, interfaces).

a

b d c b b

c

a b c

a b

b

a

a b

) where P is the Definition 1. A topology over N is a pair T = (P, ⊆ P ×N ×N ×P is the edge relation. nonempty finite set of processes and , which means that the a-interface of We write p a b q for (p, a, b, q) ∈ p is connected the b-interface of q. We require that, whenever p a b q, the following hold: (a) p = q (there are no self loops), (b) q b a p (adjacent processes are mutually connected), and   (c) for all a , b ∈ N and q  ∈ P such that p a b q  , we have a = a iff q = q  (an interface is connected to at most one process, and two distinct interfaces are connected to distinct processes). Message Sequence Charts. The possible behaviors of our systems are depicted as message sequence charts. A message sequence chart consists of a set of processes, and, for each process, of a sequence of events. Each event corresponds to an action of the process (sending or receiving a message through a given interface), and matching send and receive events are connected by a mesdef

sage edge. Events are labeled with elements of Σ = {a? | a ∈ N } ∪ {a! | a ∈ N }, according to the type of action they execute. Definition 2. A pre-message sequence chart (pre-MSC) over the set of interfaces N is a tuple M = (P, E, →, , π, λ), where – P and E are nonempty finite sets of processes and events, respectively. – π : E → P is a surjective map determining the location of an event. For def p ∈ P , we let Ep = {e ∈ E | π(e) = p}. – λ : E → Σ associates with each event the type of action that it executes. We def def let E? = {e ∈ E | ∃a ∈ N .λ(e) = a?}, and E! = {e ∈ E | ∃a ∈ N .λ(e) = a!}.

Verification of Parameterized Communicating Automata via Split-Width p1

p2

a!

b!

a!

b?

a?

b?

201

p3

p1

a!

b?

a?

a?

b!

a!

Fig. 1. An MSC with 3 processes

p2

p1

b?

a!

b!

a!

p2

p3

b? c?

Fig. 2. pre-MSCs that are not MSCs

 – → is a union p∈P →p , where each →p ⊆ Ep × Ep is the direct-successor relation of some total order on Ep . –  ⊆ E! × E? defines a bijection from E! to E? . Moreover, for each (e, f ) ∈ , π(e) = π(f ). def

Given such a pre-MSC, we define TM = (P, {(p, a, b, q) ∈ P × N 2 × P | ∃(e, f ) ∈ Ep ×Eq . (ef ∧λ(e) = a!∧λ(f ) = b?) or (f e∧λ(e) = a?∧λ(f ) = b!)}). Not all pre-MSCs correspond to actual behaviors of systems (Fig. 2). To define MSCs, we additionally require that the events are coherently ordered, and that communications are compatible with some topology and follow a FIFO policy. Definition 3. A message sequence chart (MSC) over N is a pre-MSC M = def

(P, E, →, , π, λ) such that the relation ≤ = (→ ∪ )∗ is a partial order, and: – TM as defined above is a topology, called the observable topology of M . – For all e1  e2 , f1  f2 s.t. π(ei ) = π(fi ), we have e1 ≤ f1 iff e2 ≤ f2 (FIFO). An MSC M is called compatible with a topology T when TM is a subgraph of T . Intuitively, an MSC is compatible with a topology T when it can be interpreted as a behavior over T , in which some processes may be inactive or may not use all their interfaces. We denote by MSC the set of all MSCs over N , and by MSCT the set of all MSCs compatible with a topology T . Example 4. An example MSC is depicted in Fig. 1. The vertical lines represent the succession of events on a given process, and -edges are depicted by arrows. Parameterized Communicating Automata. The idea is that each process of a topology runs one and the same automaton, whose transitions are labeled with actions of the form a!m, which emits a message m through interface a, or a?m, which receives m from interface a. The acceptance condition of a parameterized communicating automaton is given as a boolean combination of conditions of the form “at least n processes end in state s”, written #(s) ≥ n . Definition 5. A parameterized communicating automaton (PCA) over N is a tuple S = (S, ι, Msg, Δ, F ) where S is the finite set of states, ι ∈ S is the initial state, Msg is a nonempty finite set of messages, Δ ⊆ S × (Σ × Msg) × S is the transition relation, and F is the acceptance condition, a finite boolean combination of statements of the form #(s) ≥ n , with s ∈ S and n ∈ N.

202

M. Fortin and P. Gastin

The size |F | of the acceptance condition of S is defined as the length of its encoding, where all integer values are written in binary. Let M = (P, E, →, , π, λ) be an MSC. A run of S on M will be a mapping ρ : E → S satisfying some requirements. Intuitively, ρ(e) is the local state of π(e) after executing e. To determine when ρ is a run, we define another mapping, ρ− : E → S, denoting the source state of a transition: whenever f → e, we let ρ− (e) = ρ(f ); moreover, if e is →-minimal, we let ρ− (e) = ι. With this, we say that ρ is a run of S on M if, for all (e, f ) ∈ , there is a message m ∈ Msg such that (ρ− (e), (λ(e), m), ρ(e)) ∈ Δ, and (ρ− (f ), (λ(f ), m), ρ(f )) ∈ Δ. A run ρ is accepting if it satisfies the acceptance condition. In particular, ρ satisfies

#(s) ≥ n when |{e ∈ E | e is →-maximal and ρ(e) = s}| ≥ n. The set of MSCs that allow for an accepting run is denoted by L(S). Given a topology T , we let LT (S) = L(S) ∩ MSCT . Remark 6. We could add labels from a finite alphabet to the events of MSCs and to the transitions of PCAs. Such labels can be handled similarly to the λ-labeling, and all our results can easily be adapted to this setting. Similarly, allowing internal transitions for PCAs would not add any technical difficulties. Verification Problems. The non-emptiness problem asks, given a PCA S, whether its language L(S) is non empty; or in other words, whether LT (S) = ∅ for some topology T . The (local) reachability problem asks, given a PCA S and a state s of S, whether there exists a run of S in which some process reaches the state s. It can be seen as a special instance of the non-emptiness problem, by modifying the acceptance condition of S to #(s) ≥ 1 . Notice that the non-emptiness problem LT (S) = ∅ for a fixed topology T is already undecidable, since two finite automata connected by two queues can easily simulate a Turing machine. Furthermore, many decidable restrictions over fixed topologies remain undecidable in the parameterized case: for instance, bounding the number of contexts, or even of actions, performed by each process, or imposing rendez-vous synchronization (even when restricted to pipeline topologies [7]). The idea is that the unbounded number of processes can be used to construct a PCA whose behaviors are grid-like MSCs of arbitrary height and width (see Figs. 3 and 4). It is then easy to encode runs of a Turing machine: the unbounded horizontal direction encodes the tape of the machine, and the vertical direction its evolution with time. In the remaining of the paper, we will study several decidable underapproximations of the problem. For a family (Ci )i of classes of MSCs (for the concrete families studied in Sect. 4, the index i is a tuple of integers), we define the problem C-NonEmptiness as follows (and similarly, C-Reachability): Input: Question:

i in unary, a set of interfaces N , a PCA S over N L(S) ∩ Ci = ∅ ?

Verification of Parameterized Communicating Automata via Split-Width

Fig. 3. Undecidability with pipelines and rendez-vous synchronization

3

203

Fig. 4. Undecidability with a bounded number of actions on each process, using grid topologies

Split-Width

In this section, we introduce the notion of split-width, and state our decidability result for MSCs of bounded split-width. The main motivation behind split-width is that it allows to design generic decision procedures that apply to many underapproximation classes, instead of having to develop for each class a specific decision procedure with its complexity. Several examples of classes of MSCs that are captured with split-width will be given in Sect. 4. The idea of split-width is to decompose an MSC into atomic pieces consisting of a pair of matching send and receive events, using two operations: split (removing some process edges of the MSC), and divide (separating the resulting graph into two independent parts). This is described below as a two-player game. First, we introduce the notion of split MSC (an MSC missing process edges). Definition 7. A split (pre-)MSC is a tuple M = (P, O, E, →, , , π, λ), where (P, E, → ∪ , , π, λ) is a (pre-)MSC,  ∩ → = ∅, O ⊆ P is the set of open processes of M , and every split process is open, i.e., {p ∈ P | (Ep )2 ∩  = ∅} ⊆ O. The  edges are called elastic, and the → edges rigid. Processes of P \O are called closed. The intuition is that an open process of M is a process that may be missing some of its events or process edges, and an elastic edge represents a missing part of a process between two events. Any MSC can be seen as a split MSC, by taking O and  empty. A block of a split (pre-)MSC M is a maximal connected component of (E, →) on some open process. In particular, M has exactly |O| + || blocks. Several split MSCs are depicted in Fig. 5. Elastic edges are represented by red dotted lines. Open processes are indicated by dotted lines at their extremities. For instance, M  has one open process with 3 blocks, and M2 has 2 open processes, with resp. 1 and 2 blocks. We call splitting an edge of M the action of making elastic some rigid edge e → f of M . The resulting split MSC is M  = (P, O∪{π(e)}, E, →\{(e, f )}, ∪

204

M. Fortin and P. Gastin a! a?

M

b! b? a!

M

b?

a! a?

b! a! b?

a! a?

b! b? a!

b?

a! a?

b! a! b?

a! a? a! a?

b! b? b! b?

a! a? M1 a! a?

b! b? b! b?

M1

b?

b?

a!

b? a!

a!

b? b?

a!

M4

M3

a! a?

b! b?

M3

a! a?

b! b?

a! a?

b?

a!

.. .

(n ) -div (B Br B )

M2

b? M7

(n1 ) -split (ε, B1 B1 ) (n1 ) -div (B Br , B Br ) (n3 ) -split (B2 , B2 )

.. .

(n3 ) -div (B Br , Br B ) b!

b?

a!

b?

(n) -split (ε)

M2

M5

Fig. 5. A split decomposition of width 4 (Color figure online)

-split (n2 ) (B1 , ε) -div (n2 ) (B , B Br ) (n6 ) (n7 ) (B! , B? ) (B! ) (f ) a! b?

a! b?

(n4 ) (n5 ) (B! , B? ) (B? , B! ) a! b? (e) a! b?

Fig. 6. 4-DST associated with the decomposition of Fig. 5 (Color figure online)

{(e, f )}, , π, λ). For instance in Fig. 5, M  is obtained by splitting two process edges of M . We say that M can be divided into M1 = (P1 , O1 , E1 , →1 , 1 , π1 , λ1 ) and M2 = (P2 , O2 , E2 , →2 , 2 , π2 , λ2 ) when M1 and M2 are split (pre)-MSCs, and E = E1  E2 , → = →1  →2 ,  = 1  2 , π = π1  π2 , λ = λ1  λ2 , and for i ∈ {1, 2}, Oi = O ∩ Pi and i ⊆ (→ ∪ )+ . For instance, in Fig. 5, M  can be divided into M1 and M2 . A connected component of M is a split MSC M1 = (P1 , O1 , E1 , →1 , 1 , π1 , λ1 ) such that E1 is a connected component of (E, → ∪ ). Then, either M = M1 or M can be divided into M1 and some M2 . Split-Game. Let M be a split MSC with at most k blocks. A split-game with budget k on M is a two player game in which the existential player (Eve) tries to prove that M has split-width at most k, while the universal player (Adam)

Verification of Parameterized Communicating Automata via Split-Width

205

tries to disprove it. Eve begins by trying to disconnect M by splitting some of its process edges, with the condition that the resulting split MSC M  has at most k blocks. Adam then chooses a connected component M  of M  , and the game resumes on M  . Eve wins a play if it ends in an atomic MSC, i.e. a pair of matching send and receive events. She loses if she cannot disconnect a non-atomic MSC without introducing more than k blocks. The split-width of an MSC M is the minimal k such that Eve wins the splitgame with budget k on M . It is defined identically for pre-MSCs. We denote by SWk the set of MSCs of split-width at most k. Example 8. Eve wins the split-game with budget 4 on M (see Fig. 5). She starts by splitting two process edges of M , which results in the split MSC M  with three blocks. M  has two connected components, M1 and M2 , providing two choices for Adam. If he chooses M2 , Eve wins by cutting the only remaining process edge: both connected components of the resulting M2 are atomic. If he chooses M1 , Eve split the middle process edge on the first process, which creates two more blocks, and results in a total of four blocks. M1 has two isomorphic connected components, so Adam’s choices are all equivalent. Eve can then cut the two remaining process edges while still respecting her budget of four blocks. Shuffle and Merge. One can also give a bottom-up description of split-width. The duals of the split and divide operations are called respectively merge () and shuffle (). (M ) is the set of split MSCs that can be obtained by making some elastic edges of M rigid, and/or closing some of its open processes. M1 M2 is the set of split MSCs M such that M can be divided into M1 and M2 . An MSC has split-width at most k when it can be obtain by combining atomic MSCs with shuffle and merge operations, while keeping the number of blocks at most k at each step. Remark 9. The notion of open and closed processes is new. Our bound on the number of blocks (i.e. the number of open processes plus the number of elastic edges) replaces the bound on the number of elastic edges only that was used in [1, 2] for the split-width of MSCs over fixed architectures. When the topology T is fixed, the two definitions are equivalent since the number of open processes is already bounded by the number of processes in T . This is no longer true in the parameterized case. For instance, the families of MSCs defined in Figs. 3 and 4 can be decomposed into atomic pieces while using only two elastic edges, but this introduces an unbounded number of open processes. In fact, they embed a grid, hence they should have unbounded width. Remark 10. For MSCs, split-width is equivalent to tree-width and clique-width: there are linear bounds between the three measures (the proof is an easy adaptation from the non-parameterized case [10]). The motivation for introducing split-width rather than using existing measures on graphs is that it allows to take into account the specificities of MSCs, and is thus both simpler to understand and to use. In particular, using tree-width or clique-width would result in more involved automata constructions in Sect. 5.

206

M. Fortin and P. Gastin

Notice also that a bound on the split-width of an MSC M induces a bound on the tree-width of the observable topology TM (See Theorem 20). Decidability. The non-emptiness problem becomes decidable when restricted to MSCs of bounded split-width. Roughly, the proof goes as follows. First, we show that trees representing Eve’s winning strategies can be abstracted by trees over a finite alphabet (that depends only on the bound k on split-width). Then, we reduce the verification problems for PCAs to emptiness problems on tree automata. The details for these constructions will be given in Sect. 5. The proof is inspired from the non-parameterized case [2], and we show that the complexity remains the same in our setting. Theorem 11. SW-NonEmptiness is Exptime-complete, and only polynomial in the number of states and transitions of the input PCA.

4

Classes of Bounded Split-Width

The decision procedures based on split-width are generic and apply to various classes of MSCs (the main condition being that MSCs in the class have bounded split-width). When the topology is fixed, this covers many well-studied restrictions [10]. In this section, we give two examples of such classes that can be generalized to the parameterized setting: existentially bounded MSCs, and contextbounded MSCs. We also define a further extension of context-bounded MSCs, called tile-bounded MSCs, and show that it is equivalent to bounded split-width. Existentially Bounded MSCs. M is called existentially k-bounded when there exists a linearization ≤lin of its events (i.e. a total order extending ≤) such that there are at most k process or message edges going out of any prefix of the linearization: for all g ∈ E,    (e, f ) ∈ E 2 | (e  f ∨ e → f ) ∧ e ≤lin g k or Ai ∈ D and |D| = k, then v|Agt|+i (s, D) = W . D = ∅ then v2·|Agt|+i (s, D) = wAi (s) = −v3·|Agt|+i (s, D); D = ∅ then v2·|Agt|+i (s, D) = W = v3·|Agt|+i (s, D);

We take I = [[1, |Agt|]] ∪ [[2 · |Agt| + 1, 3 · |Agt|]] and J = [[|Agt| + 1, 2 · |Agt|]] ∪ [[3 · |Agt| + 1, 4 · |Agt|]]. Intuitively, the components [[1, |Agt|]] are used for immunity, the components [[|Agt| + 1, 2 · |Agt|]] are used for resilience and components [[2 · |Agt| + 1, 4 · |Agt|]] are used to constrain the payoff in case of no deviation. 5.2

Correctness of the Objectives for Robustness

Let G be a concurrent game, ρ a play of D(G) and p ∈ RAgt a payoff vector. The following lemma links the weights we chose and our solution concepts. Lemma 5. Let ρ be a play, σAgt a strategy profile and p = payoff(σAgt ). – ρ satisfies objective δ(ρ) = ∅ ⇒ MPAi (ρ) = pi if, and only if, MP2·v|Agt|+i (ρ) ≥ p(Ai ) and MP3·v|Agt|+i (ρ) ≥ −p(Ai ). – If ρ is an outcome of κ(σAgt ) then ρ satisfies objective Re(k, p) if, and only if, for all agents Ai , MPv|Agt|+i (ρ) ≥ −p(Ai ).

Robust Equilibria in Mean-Payoff Games

229

– If ρ is an outcome of κ(σAgt ), then ρ satisfies objective I(t, r, p) if, and only if, for all agents Ai , MPvi (ρ) ≥ p(Ai ) − r. – If ρ is an outcome of κ(σAgt ) with payoff(σAgt ) = p, then play ρ satisfies objective R(k, t, r, p) if, and only if, for all agents Ai , MPvi (ρ) ≥ p(Ai ) − r and MPv|Agt|+i (ρ) ≥ −p(Ai ). Putting together this lemma and the correspondence between the deviator game and robust equilibria of Theorem 6 we obtain the following proposition. Lemma 6. Let G be a concurrent game with mean-payoff objectives. There is a (k, t, r)-robust equilibrium in G if, and only if, for the multidimensional meanpayoff objective given by v, I = [[1, |Agt|]] ∪ [[2 · |Agt| + 1, 3 · |Agt|]] and J = [[|Agt| + 1, 2 · |Agt|]] ∪ [[3 · |Agt| + 1, 4 · |Agt|]], there is a payoff vector p such that Eve can ensure threshold u in D(G), where for all i ∈ [[1, |Agt|]], ui = p(Ai ) − r, u|Agt|+i = −p(Ai ), u2·|Agt|+i = p(Ai ), and u3·|Agt|+i = −p(Ai ). 5.3

Formulation of the Robustness Problem as a Polyhedron Value Problem

From the previous lemma, we can deduce an algorithm which works by querying the polyhedron value problem. Given a game G and parameters k, t, r, we ask whether there exists a payoff u that Eve can ensured in the game D(G) with multidimensional mean-payoff objective given by v, I, J, and such that for all i ∈ [[1, |Agt|]], ui + r = −u|Agt|+i = u2·|Agt|+i = −u3·|Agt|+i . As we will show in Theorem 7 thanks to Lemma 6, the answer to this question is yes if, and only if, there is a (k, t, r)-robust equilibrium. From the point of view of complexity, however, the deviator game on which we perform the query can be of exponential size compared to the original game. To describe more precisely the complexity of the problem, we remark by applying the bound of [8, Theorem 22], that given a query, we can find solutions which have a small representation. Lemma 7. If there is a solution to the polyhedron value problem in D(G) then there is one whose encoding is of polynomial size with respect to G and the polyhedron given as input. We can therefore enumerate all possible solutions in polynomial space. To obtain an polynomial space algorithm, we must be able to check one solution in polynomial space as well. This account to show that queries for the value problem in the deviator game can be done in space polynomial with respect to the original game. This is the goal of the next section, and it is done by considering small parts of the deviator game called fixed coalition games.

6

Fixed Coalition Game

Although the deviator game may be of exponential size, it presents a particular structure. As the set of deviators only increases during any run, the game

230

R. Brenguier

can be seen as the product of the original game with a directed acyclic graph (DAG). The nodes of this DAG correspond to possible sets of deviators, it is of exponential size but polynomial degree and depth. We exploit this structure to obtain a polynomial space algorithm for the value problem and thus also for the polyhedron value problem and the robustness problem. The idea is to compute winning states in one component at a time, and to recursively call the procedure for states that are successors of the current component. We will therefore consider one different game for each component. We now present the details of the procedure. For a fixed set of deviator D, the possible successors of states of the component Stat × D are the states in: Succ(D) = {TabD ((s, D), (mAgt , mAgt )) | s ∈ Stat, mAgt , mAgt ∈ Mov(s)} \ Stat× D}. Note that the size of Succ(D) is bounded by |Stat| × |Tab|, hence it is polynomial. Let u be a payoff threshold, we want to know whether Eve can ensure u in D(G), for the multi-dimensional objective defined in Sect. 5.1. A winning path ρ from a state in Stat × D is either: (1) such that δ(ρ) = D; (2) or it reaches a state in Succ(D) and follow a winning path from there. Assume we have computed all the states in Succ(D) that are winning. We can stop the game as soon as Succ(D) is reached, and declare Eve the winner if the state that is reached is a winning state of D(G). This process can be seen as a game F(D, u), called the fixed coalition game. In this game the states are those of (Stat × D) ∪ Succ(D); transitions are the same than in D(G) on the states of Stat × D and the states of Succ(D) have only self loops. The winning condition is identical to R(k, t, p) for the plays that never leave (Stat × D); and for a play that reach some (s , D ) ∈ Succ(D), it is considered winning Eve has a winning strategy from (s , D ) in D(G) and losing otherwise. In the fixed coalition game, we keep the weights previously defined for states of Stat × D, and fix it for the states that are not in the same D component by giving a high payoff on states that are winning and a low one on the losing ones. Formally, we define a multidimensional weight function v f on F(D, u) by: 1. for all s ∈ Stat, and all i ∈ [[1, 4 · |Agt|]], vif (s, D) = vi (s, D). 2. if (s, D ) ∈ Succ(D) and Eve can ensure u from (s, D ), then for all i ∈ [[1, 4 · |Agt|]], vif (s, D ) = W . 3. if (s, D ) ∈ Succ(D) and Eve cannot ensure u from (s, D ), then for all i ∈ [[1, 4 · |Agt|]], vif (s, D ) = −W − 1. Lemma 8. Eve can ensure payoff u ∈ [[−W, W ]]d in D(G) from (s, D) if, and only if, she can ensure u in the fixed coalition game F(D, p) from (s, D). Using this correspondence, we deduce a polynomial space algorithm to check that Eve can ensure a given value in the deviator game and thus a polynomial space algorithm for the robustness problem. Theorem 7. There is a polynomial space algorithm, that given a concurrent game G, tells if there is a (k, t, r)-robust equilibrium. Proof. We first show that there is a polynomial space algorithm to solve the value problem in D(G). We consider a threshold u and a state (s, D). In the fixed

Robust Equilibria in Mean-Payoff Games

231

coalition game F(D, u), for each (s , D ) ∈ Succ(D), we can compute whether it is winning by recursive calls. Once the weights for all (s , D ) ∈ Succ(D) have been computed for F(D, u), we can solve the value problem in F(D, u). Thanks to Lemma 8 the answer to value problem in this game is yes exactly when Eve can ensure u from (s, D) in D(G). There is a coNP algorithm [20] to check the value problem in a given game, and therefore there also is an algorithm which uses polynomial space. The size of the stack of recursive calls is bounded by |Agt|, so the global algorithm uses polynomial space. We now use this to show that there is a polynomial space algorithm for the polyhedron value problem in D(G). We showed in Lemma 7, that if the polyhedron value problem has a solution then there is a threshold u of polynomial size that is witness of this property. We can enumerate all the thresholds that satisfy the size bound in polynomial space. We can then test that these thresholds satisfy the given linear inequations, and that the algorithm for the value problem answers yes on this input, in polynomial space thanks to the previous algorithm. If this is the case for one of the thresholds, then we answer yes for the polyhedron value problem. The correctness of this procedure holds thanks to Lemma 7. We now use this to show that there is a polynomial space algorithm for the robustness problem. Given a game G and parameters (k, t, r), we define a tuple of linear equations, for all i ∈ [[1, |Agt|]], x2·|Agt|+i = xi + r ∧ x2·|Agt|+i = −x|Agt|+i ∧ x2·|Agt|+i = −x3·|Agt|+i (each equation can be expressed by two inequations). Thanks to Lemma 6, there is a payoff which satisfies these constraints and which Eve can ensure in D(G) if, and only if, there is a (k, t, r)-robust equilibrium. Then, querying the algorithm we described for the polyhedron value problem in D(G) with our system of inequations, answers the robustness problem.

7

Hardness

In this section, we show a matching lower bound for the resilience problem. The lower bound holds for weights that are 0 in every states except on some terminal states where they can be 1. This is also called simple reachability objectives. Theorem 8. The robustness problem is PSPACE-complete. Note that we already proved PSPACE-membership in Theorem 7. We give the construction and intuition of the reduction to show hardness and leave the proof of correctness in the appendix. We encode QSAT formulas with n variable into a game with 2 · n + 2 players, such that the formula is valid if, and only if, there is n-resilient equilibria. We assume that we are given a formula of the form φ = ∀x1 .∃x2 . ∀x3 ·∃xn . C1 ∧· · ·∧Ck , where each Ck is of the form 1,k ∨2,k ∨3,k and each j,k is a literal (i.e. xm or ¬xm for some m). We define the game Gφ as illustrated by an example in Fig. 4. It has a player Am for each positive literal xm , and a player Bm for each negative literal ¬xm . We add two extra players Eve and Adam. Eve is making choices for the existential quantification and Adam for the universal ones. When they chose a literal, the corresponding player can either go to a sink state ⊥ or continue the game to the next quantification. Once

232

R. Brenguier

a literal has been chosen for all the variables, Eve needs to chose a literal for each clause. The objective for Eve and the literal players is to reach ⊥. The objective for Adam is to reach . We ask whether there is a (n + 1)-resilient equilibrium. To a history h = Adam1 ·X1 ·Eve2 ·X2 ·Adam3 · · · Evem ·Xm with Xi ∈ {Ai , Bi }, we associate a valuation vh , such that vh (xi ) = true if Xi = Bi and vh (xi ) = false if Xi = Ai . Intuitively, Eve has to find a valuation that makes the formula hold, while Adam tries to falsify it. A1 Adam



A2 Eve

B1

⊥ B2

A3 Adam

⊥ B3

A4 Eve

⊥ B4

A1 Eve

A2 B3

B2 Eve

A3



A4

Fig. 4. Encoding of a formula φ = ∀x1 .∃x2 . ∀x3 . ∃x4 . (x1 ∨ x2 ∨ ¬x3 ) ∧ (¬x2 ∨ x3 ∨ x4 ). The dashed edges represent the strategies in the equilibrium of the players other than Eve.

References 1. Abraham, I., Dolev, D., Gonen, R., Halpern, J.: Distributed computing meets game theory: robust mechanisms for rational secret sharing and multiparty computation. In: Proceedings of the Twenty-Fifth Annual ACM Symposium on Principles of Distributed Computing, pp. 53–62. ACM (2006) 2. Alur, R., Henzinger, T.A., Kupferman, O.: Alternating-time temporal logic. J. ACM (JACM) 49(5), 672–713 (2002) 3. Aumann, R.: Acceptable points in general cooperative n-person games. Top. Math. Econ. Game Theory Essays Honor Robert J Aumann 23, 287–324 (1959) 4. Bouyer, P., Brenguier, R., Markey, N., Ummels, M.: Concurrent games with ordered objectives. In: Birkedal, L. (ed.) FOSSACS 2012. LNCS, vol. 7213, pp. 301–315. Springer, Heidelberg (2012) 5. Bouyer, P., Markey, N., Stan, D.: Mixed nash equilibria in concurrent terminalreward games. In: 34th International Conference on Foundation of Software Technology and Theoretical Computer Science (FSTTCS 2014). Leibniz International Proceedings in Informatics (LIPIcs), Dagstuhl, Germany, vol. 29, pp. 351–363 (2014) 6. Brenguier, R.: Nash equilibria in concurrent games: application to timed games. Ph.D. thesis, Cachan, Ecole normale sup´erieure (2012) 7. Brenguier, R.: Robust equilibria in concurrent games. CoRR, http://arxiv.org/ abs/1311.7683 (2015) 8. Brenguier, R., Raskin, J.-F.: Optimal values of multidimensional mean-payoff games (2014). https://hal.archives-ouvertes.fr/hal-00977352/ 9. Brenguier, R., Raskin, J.-F.: Pareto curves of multidimensional mean-payoff games. In: Kroening, D., P˘ as˘ areanu, C.S. (eds.) CAV 2015. LNCS, vol. 9207, pp. 251–267. Springer, Heidelberg (2015) 10. Brihaye, T., Bruy`ere, V., De Pril, J.: Equilibria in quantitative reachability games. In: Ablayev, F., Mayr, E.W. (eds.) CSR 2010. LNCS, vol. 6072, pp. 72–83. Springer, Heidelberg (2010)

Robust Equilibria in Mean-Payoff Games

233

11. Canny, J.: Some algebraic and geometric computations in pspace. In: Proceedings of the Twentieth Annual ACM Symposium on Theory of Computing, pp. 460–467. ACM (1988) 12. Chatterjee, K., Henzinger, T.A., Jurdzi´ nski, M.: Games with Secure Equilibria, . In: de Boer, F.S., Bonsangue, M.M., Graf, S., de Roever, W.-P. (eds.) FMCO 2004. LNCS, vol. 3657, pp. 141–161. Springer, Heidelberg (2005) 13. Chatterjee, K., Henzinger, T.A., Piterman, N.: Strategy logic. Inform. Comput. 208(6), 677–693 (2010) 14. Mogavero, F., Murano, A., Perelli, G., Vardi, M.Y.: What makes Atl* decidable? A decidable fragment of strategy logic. In: Koutny, M., Ulidowski, I. (eds.) CONCUR 2012. LNCS, vol. 7454, pp. 193–208. Springer, Heidelberg (2012) 15. Nash Jr., J.F.: Equilibrium points in n-person games. Proc. Nat. Acad. Sci. USA 36(1), 48–49 (1950) 16. Puterman, M.L.: Markov decision processes: Discrete stochastic dynamic programming (1994) 17. Ummels, M.: The complexity of nash equilibria in infinite multiplayer games. In: Amadio, R.M. (ed.) FOSSACS 2008. LNCS, vol. 4962, pp. 20–34. Springer, Heidelberg (2008) 18. Ummels, M., Wojtczak, D.: The complexity of nash equilibria in simple stochastic multiplayer games. In: Albers, S., Marchetti-Spaccamela, A., Matias, Y., Nikoletseas, S., Thomas, W. (eds.) ICALP 2009, Part II. LNCS, vol. 5556, pp. 297–308. Springer, Heidelberg (2009) 19. Ummels, M., Wojtczak, D.: The Complexity of Nash Equilibria in Limit-Average Games. In: Katoen, J.-P., K¨ onig, B. (eds.) CONCUR 2011. LNCS, vol. 6901, pp. 482–496. Springer, Heidelberg (2011) 20. Velner, Y., Chatterjee, K., Doyen, L., Henzinger, T.A., Rabinovich, A., Raskin, J.-F.: The complexity of multi-mean-payoff and multi-energy games. CoRR abs/1209.3234 (2012)

Quantifier Alternation for Infinite Words Th´eo Pierron, Thomas Place(B) , and Marc Zeitoun LaBRI, UMR 5800, University of Bordeaux, 33400 Talence, France [email protected]

Abstract. We investigate the expressive power of the quantifier alternation hierarchy of first-order logic over words. This hierarchy includes the classes Σi (sentences having at most i blocks of quantifiers starting with an ∃) and BΣi (Boolean combinations of Σi sentences). So far, this expressive power has been effectively characterized for the lower levels only. Recently, a breakthrough was made over finite words, and decidable characterizations were obtained for BΣ2 and Σ3 , by relying on a decision problem called separation, and solving it for Σ2 . The contribution of this paper is a generalization of these results to the setting of infinite words: we solve separation for Σ2 and Σ3 , and obtain decidable characterizations of BΣ2 and Σ3 as consequences.

Regular word languages form a robust class, as they can be defined either by operational, algebraic, or logical means: they are exactly those that can be defined equivalently by finite state machines (operational view), morphisms into finite algebras (algebraic view) and monadic second order (“MSO”) sentences [4, 5, 8,27] (logical view). To understand the structure of this class in depth, it is natural to classify its languages according to their descriptive complexity. The problem is to determine how complicated a sentence has to be to describe a given input language. This is a decision problem parametrized by a fragment of MSO: given an input language, can it be expressed in the fragment? This problem is called membership (is the language a member of the class defined by the fragment?). The seminal result in this field is the membership algorithm for first-order logic (FO) over finite words, which is arguably the most prominent fragment of MSO. This algorithm was obtained in two steps. McNaughton and Papert [10] observed that the languages definable in FO are exactly the star-free languages: those that may be expressed by a regular expression in which complement is allowed while the Kleene star is disallowed. Furthermore, an earlier result of Sch¨ utzenberger [23] shows that star-free languages are exactly the ones whose syntactic monoid is aperiodic. The syntactic monoid is a finite algebra that can be computed from any input regular language, and aperiodicity can be formulated as an equation that has to be satisfied by all elements of this algebra. Therefore, Sch¨ utzenberger’s result makes it possible to decide whether a regular language is star-free (and therefore definable in FO by McNaughton-Papert’s result). Following this first result, the attention turned to a deeper question: given an FO-definable language, find the “simplest” FO-sentences that define it. The standard complexity measure for FO sentences is their quantifier alternation, c Springer-Verlag Berlin Heidelberg 2016  B. Jacobs and C. L¨ oding (Eds.): FOSSACS 2016, LNCS 9634, pp. 234–251, 2016. DOI: 10.1007/978-3-662-49630-5 14

Quantifier Alternation for Infinite Words

235

which counts the number of switches between blocks of ∃ and ∀ quantifiers. This measure is justified not only because it is intuitively difficult to understand a sentence with many alternations, but also because the nonelementary complexity of standard problems for FO [25] (e.g, satisfiability) is tied to quantifier alternation. In summary, we classify FO definable languages by counting the number of quantifier alternations needed to define them and we want to be able to decide the level of a given language (which amounts to solving membership for each level). This leads to define the following fragments of FO: an FO sentence is Σi if its prenex normal form has at most i blocks of ∃ or ∀ quantifiers and starts with a block of existential ones. Note that Σi is not closed under complement (the negation of a Σi sentence is called a Πi sentence). A sentence is BΣi if it is a Boolean combination of Σi sentences (cf. figure). Clearly, we have Σi ⊆ BΣi ⊆ Σi+1 , and these inclusions are known to be strict [3, 26]: Σi  BΣi  Σi+1 .

Solving membership for levels of this hierarchy is a longstanding open problem. Following Sch¨ utzenberger’s approach, it was first investigated for languages of finite words. However, the question also makes sense for more complex structures, in particular for the most natural extension: infinite words. Sch¨ utzenberger’s result was first generalized to infinite words by Perrin [11], and a suitable algebraic framework for languages of infinite words was set up by Wilke [28]. Since a regular language of infinite words is determined by regular languages of finite words, finding a membership algorithm for languages of infinite words does not usually require to start over. Instead these algorithms are obtained by building on top of the algorithms for finite words, adding new arguments, specific to infinite words. Regarding the hierarchy, membership is easily seen to be decidable for Σ1 . For BΣ1 , the classical result of Simon [24] was generalized from finite to infinite words by Perrin and Pin [12]. For finite words, membership to Σ2 is known to be decidable [1, 15], a result lifted to infinite words in [2, 7]. Following these results, the understanding of the hierarchy remained stuck for years until the framework was extended to new and more general problems than membership. Rather than asking whether a language is definable in a fragment F, these problems ask what is the best F-definable “approximation” of this language (with respect to specific criteria). The simplest example is F-separation, which takes two regular languages as input and asks whether there exists a third language definable in F that contains the first language and is disjoint from the second. Separation is more general than membership: asking whether a regular language is definable in F is the same as asking whether it can be F-separated from its (also regular) complement. A consequence is that deciding these more general problems is usually more challenging than deciding membership. However, their investigation in the setting of finite words has also been very

236

T. Pierron et al.

rewarding. A good illustration is the transfer result of [18], which states that for all i, decidability of separation for Σi entails decidability of membership for Σi+1 . Combined with an algorithm for Σ2 -separation [18], this proved that Σ3 has decidable membership. This result was strengthened in [16], which shows that Σ3 -separation is decidable as well, thus obtaining decidability of membership for Σ4 . Finally, in [18], it was shown that BΣ2 has decidable membership by using a generalization of separation for Σ2 and analyzing an algorithm solving this generalization. It remained open to know whether it was possible to generalize with the same success this new approach to the setting of infinite words. This is the investigation that we carry out in the paper. More precisely, we rely on the crucial notion of Σi -chains, designed in [18] for presenting and proving membership and separation algorithms for finite words. We generalize this concept to infinite words and successfully use it to prove that the following problems are decidable: Σ2 -separation, Σ3 -separation, and BΣ2 membership. This demonstrates that Σi chains remain a suitable framework for presenting arguments in the setting of infinite words. On the other hand, new issues specific to infinite words arise, for example, we were not able to generalize the transfer result from Σi -separation to Σi+1 -membership (as a consequence, membership for Σ4 remains open). Note also that, for each problem, we pre-compute some information by using the corresponding algorithm designed in [16,18] for finite words. This means that the involved algorithms from [16,18] are used as subroutines of our algorithms. It is worth noting that the decidability of the membership problem for BΣ2 over infinite words has been obtained independently in [9]. While the algorithm is essentially the same as our own, its proof is completely different. We now present the problems in depth in Sect. 1, and we solve them in the rest of the paper. A detailed outline is provided at the end of Sect. 1. Due to lack of space, some proofs are postponed to the full version of this paper, see [13].

1

Presentation of the Problem

In this section, we first define the quantifier alternation hierarchy of first-order logic. Then, we present the membership problem and the separation problem. 1.1

The Quantifier Alternation Hierarchy of First-Order Logic

We fix a finite alphabet A. We denote by A+ the set of all finite nonempty words, and by A∞ the set of all infinite words over A. We use the term “word” for “finite word”. We call language (resp. language of infinite words) a subset of A+ (resp. of A∞ ). If u is a word and v is a word (resp. an infinite word), we denote by uv the word (resp. the infinite word) obtained by concatenating u to the left of v. If u is a word, we denote by u∞ the infinite word uuuu · · · obtained as the infinite concatenation of u with itself. If u is a word or an infinite word, we denote by alph(u) the alphabet of u, i.e., the set of letters of u.

Quantifier Alternation for Infinite Words

237

First-Order Logic. Any word or infinite word can be viewed as a logical structure made of a linearly ordered sequence of positions (finite for words and infinite for infinite words) labeled over alphabet A. In first-order logic “FO”, one can quantify over these positions and use the following predicates. – for each a ∈ A, a unary predicate Pa selecting all positions labeled with an a. – a binary predicate’0 , we find an NWA A and a length  such that D has a synchronizing word w with cost(w) ≤ budget if and only if A has a synchronizing nested word of length at most . The intuition behind the reduction is as follows. We encode the cost of each letter a in D with the length of a particular well-matched nested word a·wa in A; as a result, runs in D will be, in a sense, simulated by runs in A. The nested word a · wa is associated with a special gadget that we insert as a part of A; we denote this gadget pay(q, a) (there is a separate copy for each q ∈ Q). The intention is that the length of a nested word read by A corresponds to the cost of some word read by D. Obviously, there will be runs of A that have structure deviating from the form a1 · wa1 · · · ak · wak ; we call such deviations cheating. We will ensure that, along runs of interest, cheating is impossible: deviating transitions will lead to another set of gadgets, denoted punish(q), q ∈ Q. When

266

D. Chistikov et al. punish(1)

call( )

in

out ret(£)

ret( ) a b 1

a, b

2

t1,a

call(£)

ret(£)

call( )

in

p2

ret( ) out ret(£)

in pay(1, a)

b

call(x) punish(2) ret(Γ) out

force

ret(£)

1 ⇒

#

p1

t1,b

a

2

pay(1, b)

a

ret(£)

out

in

call(£)

DFA D

err

b t2,b

t2,a

err

ret(£)

call(£)

call(£) out

in

in

out

ret(£) err NWA A

pay(2, a)

pay(2, b)

ret(£)

err

Fig. 3. An example of the reduction to the Short Synchronizing Nested Word. For q ∈ {1, 2}, all #-transitions from q and from all states of gadgets pay(q, a), pay(q, b), and punish(q) lead to pq . All a, b-transitions in all states are self-loops, except in states 1, 2. The NWA A has a synchronizing nested word of length 4·budget+|wpunish |+1 if and only if D has a synchronizing word with cost at most budget.

a run of A is punished, it is forced to read a very long nested word wpunish , which results in exceeding the length . On the technical level, this “forcing” means that all shorter continuations make no progress to the synchronization objective. We now show how to construct the NWA A following this intuition; a small example is shown in Fig.  3. The set of states in A is Q ∪ {force} ∪  q∈Q,a∈Σ (pay(q, a)∪{tq,a })∪ q∈Q (punish(q)∪{pq }) where Q denotes, as above, the set of states of the DFA D, and we abuse the notation by letting pay(q, a) and punish(q) refer to the sets of states of the corresponding gadgets. The set of stack symbols of A is Γ = {x, y, £, }; the input letters are Σ∪{#} where # ∈ Σ (as in Remark on page 6, all call and return positions are assumed to have “fake” input letters call and ret). Table 1 describes transitions of A. It remains to define the gadgets pay(q, a) and punish(q). Recall that they need to let through runs on nested words wa and wpunish ; deviations are considered call(£)

in

out

err

ret( )

error NWA Bk

ret(£)

pay(q, a)

in error

error ret(£)

call( )

error

out

error

err or

NWA Bm punish(q)

Fig. 4. Gadgets pay(q, a) (on the left) and punish (on the right) where Bk , Bm are described in Example 2 with k = 4 · cost(a) − 3 and m = |wpunish | − 2

Synchronizing Automata over Nested Words

267

cheating and are handled appropriately. We base the construction of pay(q, a) and punish(q) on the family of NWA Bn from Example 2; see Fig. 4. Each gadget has two designated local states in and out, and the shortest run from in to out is over the nested word that we denote by va (where wa = call · va · ret) in pay(q, a) and by vpunish (where wpunish = call·vpunish ·ret) in punish(q). We pick the parameter k = |va | in Bk in such a way that |a · wa | = |a · call · va · ret| = 4 · cost(a); note that k = 4 · cost(a) − 3 ≥ 1, since cost(a) ≥ 1. Our choice for m in Bm will be given below. Now recall that the NWA Bn in Example 2 had only partially defined transition functions; we make them complete by directing all missing transitions (shown as “errors” in Fig. 4) to in in punish and to new local states err in pay. Note that this includes missing transitions on call (they all push x to the stack) and missing transitions on ret (at every control state, there is a popping transition for each γ ∈ Γ). In contrast, on input # all transitions from pay(q, a) and punish(q) go to the state pq . In fact, every synchronizing word is forced to have at least one occurrence of #, otherwise the run starting from yet another state force cannot be synchronized with other runs. Therefore, every synchronizing word needs to have at least one occurrence of wpunish , and this determines our choice of  and |wpunish |. It is natural to pick  = 1 + |wpunish | + 4 · budget; since we want to have  < 2 · |wpunish |, we need to make sure that |wpunish | > 4 · budget + 1. We thus choose m + 2 = |wpunish | = 4 · budget + 2 and  = 8 · budget + 3. This completes the description of our reduction; we omit the proof of correctness because of space constraints. This reduction provides the lower bound in Theorem 6. Acknowledgements. The authors are grateful to Michael Wehar for comments.

References 1. Alur, R., Madhusudan, P.: Adding nesting structure to words. J. ACM 56(3), 16 (2009) 2. Caucal, D.: Synchronization of pushdown automata. In: Ibarra, O.H., Dang, Z. (eds.) DLT 2006. LNCS, vol. 4036, pp. 120–132. Springer, Heidelberg (2006) 3. Chytil, M.P., Monien, B.: Caterpillars and context-free languages. In: Choffrut, C., Lengauer, T. (eds.) STACS 90. LNCS, vol. 415, pp. 70–81. Springer, Heidelberg (1990) 4. Comon, H., Dauchet, M., Gilleron, R., L¨ oding, C., Jacquemard, F., Lugiez, D., Tison, S., Tommasi, M.: Tree Automata Techniques and Applications, 12 October 2007. http://www.grappa.univ-lille3.fr/tata 5. Doyen, L., Juhl, L., Larsen, K.G., Markey, N., Shirmohammadi, M.: Synchronizing words for weighted and timed automata. In: 34th International Conference on Foundation of Software Technology and Theoretical Computer Science, FSTTCS, 15–17 December 2014, New Delhi, India, pp. 121–132 (2014) 6. Driscoll, E., Thakur, A., Reps, T.: OpenNWA: a nested-word automaton library. In: Madhusudan, P., Seshia, S.A. (eds.) CAV 2012. LNCS, vol. 7358, pp. 665–671. Springer, Heidelberg (2012)

268

D. Chistikov et al.

7. Eppstein, D.: Reset sequences for monotonic automata. SIAM J. Comput. 19(3), 500–510 (1990) 8. Ershov, A.P.: On programming of arithmetic operations. Commun. ACM 1(8), 3–9 (1958) 9. Esparza, J., Ganty, P., Kiefer, S., Luttenberger, M.: Parikh’s theorem: a simple and direct automaton construction. Inf. Process. Lett. 111(12), 614–619 (2011) 10. Esparza, J., Ganty, P., Majumdar, R.: Parameterized verification of asynchronous shared-memory systems. In: Sharygina, N., Veith, H. (eds.) CAV 2013. LNCS, vol. 8044, pp. 124–140. Springer, Heidelberg (2013) 11. Esparza, J., Luttenberger, M., Schlund, M.: A brief history of Strahler numbers. In: Dediu, A.-H., Mart´ın-Vide, C., Sierra-Rodr´ıguez, J.-L., Truthe, B. (eds.) LATA 2014. LNCS, vol. 8370, pp. 1–13. Springer, Heidelberg (2014) 12. Fominykh, F.M., Martyugin, P.V., Volkov, M.V.: P(l)aying for synchronization. Int. J. Found. Comput. Sci. 24(6), 765–780 (2013) 13. Hopcroft, J.E., Paul, W.J., Valiant, L.G.: On time versus space. J. ACM 24(2), 332–337 (1977) 14. Kozen, D.: Lower bounds for natural proof systems. In: 18th Annual Symposium on Foundations of Computer Science, Providence, Rhode Island, USA, 31 October– 1 November, pp. 254–266 (1977) 15. Lengauer, T., Tarjan, R.E.: The space complexity of pebble games on trees. Inf. Process. Lett. 10(4/5), 184–188 (1980) 16. Martyugin, P.: Computational complexity of certain problems related to carefully synchronizing words for partial automata and directing words for nondeterministic automata. Theor. Comput. Syst. 54(2), 293–304 (2014) 17. Mehlhorn, K.: Pebbling moutain ranges and its application of DCFL-recognition. In: Proceedings Automata, Languages and Programming, 7th Colloquium, Noordweijkerhout, The Netherland, July 14–18, pp. 422–435 (1980) 18. Olschewski, J., Ummels, M.: The complexity of finding reset words in finite automata. In: Hlinˇen´ y, P., Kuˇcera, A. (eds.) MFCS 2010. LNCS, vol. 6281, pp. 568–579. Springer, Heidelberg (2010) 19. Paterson, M.S., Hewitt, C.E.: Comparative schematology. In: Record of the Project MAC Conference on Concurrent Systems and Parallel Computation, pp. 119–127. ACM, MIT AI Memo AIM-201 (1970). http://hdl.handle.net/1721.1/5851 ´ On two combinatorial problems arising from automata theory. 20. Pin, J.-E.: North-Holland Math. Stud. 75, 535–548 (1983) 21. Rystsov, I.K.: Polynomial complete problems in automata theory. Inf. Process. Lett. 16(3), 147–151 (1983) 22. Sandberg, S.: Homing and synchronizing sequences. In: Broy, M., Jonsson, B., Katoen, J.-P., Leucker, M., Pretschner, A. (eds.) Model-Based Testing of Reactive Systems. LNCS, vol. 3472, pp. 5–33. Springer, Heidelberg (2005) 23. Savage, J.E.: Models of Computation - Exploring the Power of Computing. Addison-Wesley, Boston (1998) ˇ 24. Cern´ y, J., Pirick´ a, A., Rosenauerov´ a, B.: On directable automata. Kybernetika 7(4), 289–298 (1971) ˇ 25. Volkov, M.V.: Synchronizing automata and the Cern´ y conjecture. In: Mart´ın-Vide, C., Otto, F., Fernau, H. (eds.) LATA 2008. LNCS, vol. 5196, pp. 11–27. Springer, Heidelberg (2008) 26. Wehar, M.: Hardness results for intersection non-emptiness. In: Esparza, J., Fraigniaud, P., Husfeldt, T., Koutsoupias, E. (eds.) ICALP 2014, Part II. LNCS, vol. 8573, pp. 354–362. Springer, Heidelberg (2014)

On Freeze LTL with Ordered Attributes Normann Decker(B) and Daniel Thoma Institute for Software Engineering and Programming Languages, University of L¨ ubeck, L¨ ubeck, Germany {decker,thoma}@isp.uni-luebeck.de

Abstract. This paper is concerned with Freeze LTL, a temporal logic on data words with registers. In a (multi-attributed) data word each position carries a letter from a finite alphabet and assigns a data value to a fixed, finite set of attributes. The satisfiability problem of Freeze LTL is undecidable if more than one register is available or tuples of data values can be stored and compared arbitrarily. Starting from the decidable one-register fragment we propose an extension that allows for specifying a dependency relation on attributes. This restricts in a flexible way how collections of attribute values can be stored and compared. This conceptual dimension is orthogonal to the number of registers or the available temporal operators. The extension is strict. Admitting arbitrary dependency relations, satisfiability becomes undecidable. Tree-like relations, however, induce a family of decidable fragments escalating the ordinalindexed hierarchy of fast-growing complexity classes, a recently introduced framework for non-primitive recursive complexities. This results in completeness for the class F0 . We employ nested counter systems and show that they relate to the hierarchy in terms of the nesting depth.

1

Introduction

A central aspect in modern programming languages and software architectures is dynamic and unbounded creation of entities. In particular object oriented designs rely on instantiation of objects on demand and flexible multi-threaded execution. Finite abstractions can hardly reflect these dynamics and therefore infinite models are very valuable for specification and analysis. This motivates us to study the theoretical framework of words over infinite alphabets. It allows for abstracting, e.g., the internal structure and state of particular objects or processes while still being able to capture the architectural design in terms of interaction and relations between dynamically instantiated program parts. These data words, as we consider them here, are finite, non-empty sequences w = (a1 , d1 )(a2 , d2 ) . . . (an , dn ) where the i-th position carries a letter ai from a finite alphabet Σ. Additionally, for a fixed, finite set of attributes A a data valuation di : A → Δ assigns to each attribute a data value from an infinite domain Δ with equality. This work was partially supported by EGIDE/DAAD-Procope (FREQS). c Springer-Verlag Berlin Heidelberg 2016  B. Jacobs and C. L¨ oding (Eds.): FOSSACS 2016, LNCS 9634, pp. 269–284, 2016. DOI: 10.1007/978-3-662-49630-5 16

270

N. Decker and D. Thoma

Freeze LTL. In formal verification, temporal logics are widely used for formulating behavioural specifications and, regarding data, the concept of storing values in registers for comparison at different points in time is very natural. This paper is therefore concerned with the logic Freeze LTL [8] that extends classical Linear-time Temporal Logic (LTL) by registers and was extensively studied during the past decade. Since the satisfiability problem of Freeze LTL is undecidable in general, we specifically consider the decidable fragment LTL↓1 [7] that is restricted to a single register and future-time modalities. More precisely, we propose a generalisation of this fragment and study the consequences in terms of decidability and complexity. Considering specification and modelling, the ability of comparing tuples of data values arbitrarily is a valuable feature. Unfortunately, this generically renders logics on data words undecidable (cf. related work below). We therefore extend Freeze LTL by a mechanism for carefully restricting the collections of values that can be compared in terms of a dependency relation on attributes. In general, this does not suffice to regain decidability of the satisfiability problem. Imposing, however, a hierarchical dependency structure such that comparison of attribute values is carried out in an ordered fashion, we obtain a strict hierarchy of decidable fragments parameterised by the maximal depth of the attribute hierarchy. Before we exemplify this concept, let us introduce basic notation. Let Σ be a finite alphabet and (A, ) a finite set of attributes together with a reflexive and transitive relation  ⊆ A × A, i.e., a quasi-ordering, simply denoted A if  is understood. We call our logic LTL↓qo and define its syntax according to the grammar ϕ::=a | ¬ϕ | ϕ ∧ ϕ | X ϕ | ϕUϕ | ↓x ϕ | ↑x for letters a ∈ Σ and attributes x ∈ A. We further include common abbreviations such as disjunction, implication or the temporal operators release (ϕ R ψ := ¬(¬ϕU¬ψ)), weak next (Xϕ := ¬ X ¬ϕ) and globally (G ϕ := false R ϕ). The restriction of LTL↓qo to a particular, fixed set of attributes (A, ) is denoted LTL↓(A,) (or simply LTL↓A ). In the following, we explain the idea of our extension by means of an example. The formal semantics is defined in Sect. 2. Example 1. Consider a system with arbitrarily many processes that can lock, unlock and use an arbitrary number of resources. A data word over the alphabet Σ = {lock, unlock, use, halt} can model its behaviour in terms of an interleaving of individual actions and global signals. The corresponding data valuation can provide specific properties of an action, such as a unique identifier for the involved process and the resource. Let us use attributes A = {pid, res} and interpret data values from Δ as IDs. Notice that this way, we do not assume a bound on the number of involved entities. Consider now the property that locked resources must not be used by foreign processes and all locks must be released on system halt. To express this, we need to store both the process and resource ID for every lock action and verify that

On Freeze LTL with Ordered Attributes

(res) (pid)

lock lock use unlock unlock 1 2 2 2 1 1

1

1

1

1

271

lock lock use unlock halt unlock 1 3 3 3 9 1 1

2

1

1

8

1

Fig. 1. The left word satisfies the formula from Example 1 whereas every strict prefix does not. The right word violates the property because at position 3 use holds and the value of res matches the one stored at position 2 but the whole valuation (3,1) differs from (3,2), so the check ↑pid fails. Moreover, halt occurs before (1,1) was observed again in combination with unlock.

a use involving the same resource also involves the same process. As mentioned earlier, employing a too liberal mechanism to store multiple data values at once breaks the possibility of automatic analysis. In our case, however, we do not need to refer to processes independently. It suffices to consider only resources individually and formulate that the particular process that locks a resource is the only one using it before unlocking. This one-to-many correspondence between processes and resources allows us to declare the attribute pid to be dependent on the attribute res and formulate the property by the formula G(lock → ↓pid ((use ∧ ↑res → ↑pid ) ∧ ¬halt)U(unlock ∧ ↑pid )). The freeze quantifier ↓pid stores the current value assigned to pid and also implicitly that of all its dependencies, res in this case. The check operator ↑x , for an attribute x ∈ A, then verifies at some position that the current values of x and its dependencies coincide with the information that was stored earlier. Also, properties independent of the data can be verified within the same context, e.g., ¬halt for preventing a shut down as long as any resource is still locked. See Fig. 1 for example words. Using this extended storing mechanism, we can select the values of the two attributes (↓pid ) and identify and distinguish positions in a data word where both (↑pid ), a particular one of them (↑res ) or a global signal (e.g., halt) occurs. In contrast to other decidable fragments of Freeze LTL, we are thus able to store collections of values and can compare individual values across the hierarchy of attributes. This allows for reasoning on complex interaction of entities, also witnessed by the high, yet decidable, complexity of the logic. Outline and Results. We define the semantics of LTL↓qo in Sect. 2 generalising Freeze LTL based on quasi-ordered attribute sets. We show that every fragment LTL↓A is undecidable unless A has a tree-like structure, formalised as what we call a tree-quasi-ordering. Section 3 is devoted to nested counter systems (NCS) and an analysis of their coverability problem. We determine its non-primitive recursive complexity in terms of fast-growing complexity classes [20]. These classes Fα are indexed by ordinal numbers α and characterise complexities by fast-growing functions from the extended Grzegorczyk hierarchy (details are provided in Sect. 3). We show that with increasing nesting level coverability in NCS exceeds every class Fα for ordinals α < 0 . By also providing a matching upper bound, we establish the following.

272

N. Decker and D. Thoma

Theorem 2 (NCS). The coverability problem in NCS is F0 -complete. We consider the fragment LTL↓tqo in Sect. 4. It restricts the available dependency relations to tree-quasi-orderings. By reducing the satisfiability problem to NCS coverability, we obtain a precise characterisation of the decidability frontier in LTL↓qo . Moreover, we transfer the lower bounds obtained for NCS to the logic setting. This leads us to a strict hierarchy of decidable fragments of LTL↓tqo parameterised by the depth of the attribute orderings and a completeness result for LTL↓tqo . Theorem 3 (LTL↓qo ). The satisfiability problem of – LTL↓A is decidable if and only if A is a tree-quasi-ordering. – LTL↓tqo is F0 -complete. Related Work. The freeze [13] mechanism was introduced as a natural form of storing and comparing (real-time) data at different positions in time [1] and since studied extensively in different contexts, e.g., [10, 11, 17]. In particular linear temporal logic employing the freeze mechanism over domains with only equality, i.e., data words, was considered in [8] and shown highly undecidable (Σ11 -hard). Therefore, several decidable fragments were proposed in the literature with complexities ranging from exponential [15] and double-exponential space [6] to nonprimitive recursive complexities [7]. For the one-register fragment LTL↓1 that we build on here, an Fω upper bound was given in [9]. Due to its decidability and expressiveness, it is called in [7] a “competitor” for the two-variable firstorder logic over data words FO2 (∼, α1 ≥ . . . ≥ αk . Furthermore, we denote limit ordinals by λ. These are ordinals such that α + 1 < λ for every α < λ. We associate them with a fundamental sequence (λn )n with supremum λ defined by n

(α + ω

β+1

   )n := α + ω β + . . . + ω β

and





(α + ω λ )n := α + ω λn

for ordinals β and limit ordinals λ . Then, 0 is the smallest ordinal α such that α = ω α . We denote the n-th exponentiation of ω as Ωn , i.e. Ω1 := ω and Ωn+1 := ω Ωn . Consequently, (Ωn )m < Ωn is the m-th element of the fundamental sequence of Ωn . Given a monotone and expansive1 function h : N → N, a Hardy hierarchy is an ordinal-indexed family of functions hα : N → N defined by h0 (n) := n, hα+1 (n) := hα (h(n)) and hλ (n) := hλn (n). Choosing h as the incrementing function H(n) := n + 1, the fast growing hierarchy is the family of α functions Fα (n) with Fα (n) := H ω (n). 1

A function f : A → A over an ordering (A, ≤) is monotone if a ≤ a ⇒ f (a) ≤ f (a ) and expansive if a ≤ f (a) for all a, a ∈ A.

On Freeze LTL with Ordered Attributes

277

The hierarchy of fast growing complexity classes Fα for ordinals α is defined in terms of the fast-growing functions Fα . We refer the reader to [20] for details and only remark that F 0 in the NCS configuration is encoded by a letter (q, k − i) with priority k − i. E.g., the 3-NCS configuration q0 (q1 + q1 (q2 + q2 ) + q1 (q2 + q2 + q3 (q4 ))) can be encoded as a channel of the form (q1 , 2)(q1 , 2)(q2 , 1)(q2 , 1)(q1 , 2)(q2 , 1)(q2 , 1)(q3 , 1)(q4 , 0) while q0 is encoded in the finite control. Taking the highest priority for the outermost level ensures that the lossiness of PCS corresponds to descending with respect to  for the encoded NCS configuration. Thus the coverability problem in NCS directly translates to that in PCS. The coverability (control-state reachability) problem in PCS with one channel and k priorities lies in the class FΩ2k [12] and we thus obtain an upper bound for NCS coverability. Proposition 6. Coverability in k-NCS is in FΩ2k . Lower Bound. We can reduce, for any k > 1, the halting problem of H (Ωk )l bounded Minsky machines to coverability in k-NCS with the number of states bounded by l + c, for some constant c. This yields the following characterisation (recall that H (Ωk+1 )l = F(Ωk )l ). Theorem 7. Coverability in (k + 1)-NCS is FΩk -hard. The idea is to construct a k-NCS N = (Q, δ) that can simulate the evaluation of the Hardy function H α (n) for α ≤ (Ωk )l in forward as well as backward direction. It can then compute a budget that is used for simulating the Minsky machine. Lower bounds for various models were obtained using this scheme for Turing machines [4, 12] or Minsky machines [19, 21]. The following construction uses k + 1 levels of which one can be eliminated later. We encode the ordinal parameter α of H α (n) and its argument n ∈ N (unary) into a configuration n

Cα,n

   := main(s(Mα ) + c(1 + . . . + 1))

278

N. Decker and D. Thoma

using control-states main, s, c, ω ∈ Q and configurations Mα defined by M0 := ∅ and Mωα +β := ω(Mα ) + Mβ . For example, an ordinal α = ω ω + ω 2 + ω 2 + 1 is encoded by Mα = {(ω, {(ω, {(ω, ∅) : 1}) : 1}) : 1, (ω, {(ω, ∅) : 2}) : 2, (ω, ∅) : 1} 1

Note that we use shorthands for readability, e.g., ω ω stands for ω ω where 1 is again short for the ordinal ω 0 . The construction has to fulfil the following two properties. As NCS do not feature a zero test exact simulation cannot be enforced but errors can be restricted to be “lossy”. 

Lemma 8. For all configurations Cα,n →∗ Cα ,n we have H α (n) ≥ H α (n ). The construction will, however, admit at least one run maintaining exact values. 

Lemma 9. If H α (n) = H α (n ) then there is a run Cα,n →∗ Cα ,n . The main challenge is simulating a computation step from a limit ordinal to an element of its fundamental sequence, i.e., from Cα+λ,n to Cα+λn ,n and conversely. Encoding the ordinal parameter using multisets loses the ordering of the addends of the respective CNF terms. Thus, instead of taking the last element of the CNF term we have to select the smallest element, with respect to , of the corresponding multiset. To achieve that, we extend NCS by two operations cp and min. Given some configuration C = q1 (q2 (M )) ∈ CN the operation (q1 , q2 )cp(q1 , q2 ) copies M resulting in C  = q1 (q2 (M1 ) + q2 (M2 )) with M1 , M2  M . Conversely, given the configuration C  the operation (q1 , q2 )min(q1 , q2 ) results in C with M  M1 , M2 . Both operations can be implemented in a depth first search fashion using the standard NCS operations. Based on them the selection of a smallest element from a multiset can be simulated: all elements are copied (non-deterministically) one by one to an auxiliary set while enforcing a descending order. Applying the min operation in every step ensures that we either proceed indeed in descending order or make a “lossy” error. We guess, in each step, whether the smallest element is reached and in that case delete the source multiset. Thereby it is ensured, that the smallest element has been selected or, again, a “lossy” error occurs such that the selected element is now the smallest one. The additional level in the encoding of Cα,n enables us to perform this deletion step. A similar idea to select a smallest ordinal from a multiset is used in [19]. However, we need to handle nested structures of variable size correctly whereas in this work the considered ordinals are below Ω3 . They are represented by a multiset of vectors of fixed length where the vectors can be compared and modified directly in order enforce the choice of a minimal one. We now construct an NCS simulating an H α (s)-bounded Minsky machine M of size s := |M| analogously to the constructions in [4, 12, 21]. It starts in a configuration Cα,s to evaluate H α (s). When it reaches C0,n for some n ≤ H α (s) it switches its control state and starts to simulate M using n as a budget for the sum of the two simulated counters. Zero tests can then be simulated by

On Freeze LTL with Ordered Attributes

279

resets (deleting and creating multisets) causing a “lossy” error in case of an actually non-zero counter. When the simulation of M reaches a final state the NCS moves the current counter values back to the budged counter and performs a construction similar to the one above but now evaluating H α (s) backwards until reaching (Cα,s ) , the initial configuration with a different control state. If (Cα,s ) can be reached (or even covered) no “lossy” errors occurred and the Minsky machine M was thus simulated correctly regarding zero tests.

4

From LTL↓tqo to NCS and Back

Theorem 5 established a necessary condition for LT L↓A to have a decidable satisfiability problem, namely that A is a tree-quasi-ordering. In the following we show that this is also sufficient. Let LTL↓tqo denote the fragment of LTL↓qo restricted to tree-quasi-ordered sets of attributes. The decidability and complexity results for NCS can be transferred to LTL↓tqo to obtain upper and lower bounds for the satisfiability problem of the logic. We show a correspondence between the nesting depth in NCS and the depths of the tree-quasi-ordered attribute sets that thus constitutes a semantic hierarchy of logical fragments. We provide the essential ideas in the following. The depth of a finite tree-quasi-ordering A is the maximal length k of strictly increasing sequences x1  x2  . . .  xk of attributes in A. The first observation is that we can reduce satisfiability of any LTL↓tqo formula over attributes A to satisfiability of an LTL↓[k] formula where [k] = {1, . . . , k} is an initial segment of the natural numbers with natural linear ordering and k is the depth of A. Proposition 10 (LTL↓tqo to LTL↓[k] ). For a tree-quasi-ordered attribute set A of depth k every LTL↓A formula can be translated to an equisatisfiable LTL↓[k] formula of exponential size.

To reduce an arbitrary tree-quasi-ordering A of depth k we first remove maximal strongly connected components (SCC) in the graph of A and replace each of them by a single attribute. This does only affect the semantics of formulae ϕ if attributes are compared that did not have an isomorphic downward-closure in A. These cases can, however, be handled by additional constraints added to ϕ. Data words over a thus obtained partially ordered attribute set of depth k can now be encoded into words over the linear ordering [k] of equal depth k. The idea is to encode a single position into a frame of positions in the fashion of [5, 14]. That way a single attribute on every level suffices. Any formula can be transformed to operate on these frames instead of single positions at the cost of an at most exponential blow-up. From LTL↓[k] to NCS. Given an LTL↓[k] formula Φ we can now construct a (k+1)-NCS N and two configurations Cinit , Cf inal ∈ CN such that Φ is satisfiable if and only if Cf inal can be covered from Cinit . The idea is to encode sets of guarantees into NCS configurations. These guarantees are subformulae of Φ and are guaranteed to be satisfiable. The constructed

280

N. Decker and D. Thoma 1

2



3



{ϕ1 } 4

5

{ϕ2 , ψ2 }

{ψ3 }

6

{ϕ3 } C

9

7



8



{a}

{↓2 ϕ2 }

{ϕ1 } →∗

{¬ ↑2 } {↑1 }

{ϕ2 ∧ ψ2 } {ψ3 }

{ϕ3 , ↑3 }

{¬b} {a}

C

Fig. 2. Example of a guarantee forest of depth 3 maintained and modified by the NCS constructed for some LTL↓[3] formula. Node enumeration (grey) is only for reference.

NCS can instantiate new guarantees and combine existing ones while maintaining the invariant that there is always a data word w ∈ (Σ × Δ[k] )+ that satisfies all of them. To ensure the invariant, the guarantees are organised in a forest of depth k as depicted in Fig. 2. All formulae ϕ contained in the same node v of this forest are moreover not only satisfied by the same word w but also with respect to a common valuation [k] dv ∈ Δ⊥ , i.e., (w, 1, dv ) |= ϕ. Recall that valuations over linearly ordered attributes can be seen as sequences. The forest structure now represents the common-prefix relation between these valuations dv . For two nodes v, v  having a common ancestor at level i ∈ [k] in the forest, the corresponding valuations dv , dv can be chosen such that they agree on attributes 1 to i. A uniquely marked branch in the forest further represents the valuation d1 at the first position in w. If a formula ϕ is contained in the marked node at level i in the forest then [k] (w, 1, d1 |i ) |= ϕ. In that case (w, 1, d) |= ↓i ϕ holds for any d ∈ Δ⊥ and the formula ↓i ϕ could be added to any of the nodes in the forest without violating the invariant. Similarly, for a marked node v at level i the formulae ↑i can be added to any node in the subtree with root v. Moreover, other atomic formulae, Boolean combinations, and temporal operators can also be added consistently. The NCS N can perform such modifications on the forest, represented by its configuration, by corresponding transitions. Example 11. Consider the two guarantee forests depicted in Fig. 2 that are encoded in configurations C and C  of an NCS constructed for some LTL↓[3] formula. The invariant is the existence of a word w = (a, d) . . . and valuations dv ∈ Δ[i] such that (w, 1, dv ) satisfies the formulae in a node v at level i. The forest structure relates these valuations to d (nodes marked by ✓) and each other. E.g., (w, 1, d|1 ) |= ϕ1 , (w, 1, d|2 ) |= ϕ2 and there is e with e|2 = d|2 and e(3) = d(3) s.t. (w, 1, e) |= ψ3 . Let v1 , . . . , v9 be the nodes of the forest (as enumerated in the figure). Several possible operations are exemplified by the transition between C and C  . The formula ↑3 can be added to the node v6 containing the formula ϕ3 since that node is checked on level 3. Similarly, there is d3 for node v3 such that d3 (1) = d(1) and hence (w, d3 ) |= ↑1 . The formula ↑2 cannot be added to the node v2 since it is not below the checked node on level two. Consequently, the node can contain ¬↑2 . Node v4 on level

On Freeze LTL with Ordered Attributes

281

2 does already contain ϕ2 and ψ2 , meaning they are both satisfied by w and a valuation d4 ∈ Δ[2] . Hence the same holds for their conjunction. Moreover, v4 is checked and therefore d4 = d|2 . This implies that (w, d ) |= ↓2 ϕ2 for any d and that the formula can be added to any node in the tree, e.g. v7 . Recall that we only need to consider subformulae of Φ and thus remain finitestate for representing nodes. More precisely, the number of states in N is exponential in the size of Φ since they encode sets of formulae. A crucial aspect is how the NCS can consistently add formulae of the form X ϕ. This needs to be done for all stored guarantees at once but NCS do not have an atomic operation for modifying all states in a configuration. Therefore, the forest is copied recursively, processing each copied node. The NCS N can choose at any time to stop and remove the remaining nodes. That way it might loose guarantees but maintains the invariant since only processed nodes remain in the configuration. The forest of depth k itself could be maintained by a k-NCS but to implement the copy operation an additional level is needed. The initial configuration Cinit consists of a forest without any guarantees. In a setup phase, the NCS can add branches and formulae of the form Xϕ since they are all satisfied by any word of length 1. Once the formula Φ is encountered in the current forest the NCS can enter a specific target state qf inal . A path starting in Cinit and covering the configuration Cf inal = qf inal then constitutes a model of Φ and vice versa. Theorem 12. For tree-quasi-ordered attribute sets A with depth k satisfiability of LTL↓A can be reduced in exponential space to coverability in (k + 1)-NCS. From k-NCS to LTL↓[k] . Let N = (Q, δ) be a k-NCS. We are interested in describing witnesses for coverability. It suffices to construct a formula ΦN that characterises precisely those words that encode a lossy run from a configuration Cstart to a configuration Cend . We call a sequence C0 C1 . . . Cn of configurations Cj ∈ CN a lossy run from C0 to Cn if there is a sequence of intermediate  such that Ci  Ci → Ci+1 for 0 ≤ i < n. Then Cend configurations C0 . . . Cn−1 is coverable from Cstart if and only if there is a lossy run from Cstart to some Cn  Cend . A configuration of a k-NCS is essentially a tree of depth k + 1 and can be encoded into a [k]-attributed data word as a frame of positions, similar as done to prove Proposition 10. We use an alphabet Σ where every letter a ∈ Σ encodes, among other information, a (k+1)-tuple of states from Q, i.e., a possible branch in the tree. Then a sequence of such letters represents a set of branches that form a tree. The data valuations represent the information which of the branches share a common prefix. Further, this representation is interlaced: it only uses odd positions. The even position in between are used to represent an exact copy of the structure but with distinct data values. We use appropriate LTL↓[k] formulae to express this shape. Figure 3 shows an example. To be able to formulate the effect of transition rules without using pasttime operators we encode lossy runs reversed. Given that a data word encodes

282

N. Decker and D. Thoma q0 q1 q2

q1 q3

q4 q5

q0 q0 q0 q 0 q 0 q 0 q 0 q0 q1 q1 q1 q 1 q 1 q 1 q 4 q4 q2 q2 q3 q3 − − q5 q5 1 10 1 10 6 60 4 40 2 20 3 30 7 70 5 50

Fig. 3. Encoding of a 2-NCS configuration (l.) as [2]-attributed data word (r.). Instead of letters from Σ the encoded tuples of states from Q are displayed at every position.

a sequence C0 C1 . . . Cn of configurations as above we model the (reversed) control flow of the NCS N = (Q, δ) by requiring that every configuration but for the last be annotated by some transition rule tj ∈ δ for 0 ≤ j < n. The labelling is encoded into the letters from Σ and we impose that this transition sequence actually represents the reversal of a lossy run. That is, for every configuration Cj in the sequence (for 0 ≤ j < n) with annotated transition rule tj there is a configuration tj

  (not necessarily in the sequence) such that Cj ← Cj+1  Cj+1 . Cj+1 For the transition tj to be executed correctly (up to lossiness) we impose that every branch in Cj must have a corresponding branch in Cj+1 . Yet, there may be branches in Cj+1 that have no counterpart in Cj and were thus lost upon executing tj . Shared data values are now used to establish a link between corresponding branches: for every even position in the frame that encodes Cj there must be an odd position in the consecutive frame (thus encoding Cj+1 ) with the same data valuation. To ensure that links are unambiguous we require that every data valuation occurs at most twice in the whole word. Depending on the effect of the current transition the letters of linked positions are related accordingly. E.g., for branches not affected at all by tj the letters are enforced to be equal. This creates a chain of branches along the run that are identified: an odd position links forward to an even one, the consecutive odd position mimics it and links again forward. Based on these ideas we can construct a formula satisfied precisely by words encoding a lossy run between particular configurations. The size of the formula is polynomial in the size of the NCS N and can be built by instantiating a set of patterns while iterating over the transitions and states of N , requiring logarithmic space to control the iterations.

Theorem 13. The coverability problem of k-NCS can be reduced in logarithmic space to LTL↓[k] satisfiability.

5

Conclusion

By Theorem 12 together with Proposition 6 and Theorem 13 with Theorem 7 we can now characterise the complexity of LTL↓tqo fragments as follows. Proposition 14. Satisfiability of LTL↓A over a tree-quasi-ordered attribute set of depth k is in FΩ2(k+1) and FΩk -hard.

On Freeze LTL with Ordered Attributes

283

Together with Theorem 5 this completes the proof of Theorem 3 stating that LTL↓tqo is the maximal decidable fragment of LTL↓qo and F0 -complete. The result also shows that the complexity of the logic continues to increase strictly with the depth of the attribute ordering. The logics ND-LTL± were shown to be decidable by reduction to NCS [5]. Our results thus provide a first upper bound for their satisfiability problem. Moreover, we derive significantly improved lower bounds by applying the construction to prove Theorem 13 analogously to ND-LTL+ and, with reversed encoding, to the past fragment ND-LTL− . A subtle difference is that an additional attribute level is needed in order to express the global data-aware navigation needed to enforce the links between encoded configurations. Corollary 15. Satisfiability of ND-LTL± with k + 1 levels is in FΩ2(k+1) and FΩk -hard. PCS were proposed as a “master problem” for F0 [12] and indeed our upper complexity bounds for NCS rely on them. However, they are not well suited to prove our hardness results. This is due to PCS being based on sequences and the embedding ordering while NCS are only based on multisets and the subset ordering. In a sense, PCS generalise the concept of channels to multiple levels of nesting, whereas NCS generalise the concept of counters. Hence, we believe NCS are a valuable addition to the list of F0 -complete models. They may serve well to prove lower bounds for formalisms that are like Freeze LTL more closely related to the concept of counting.

References 1. Alur, R., Henzinger, T.A.: A really temporal logic. J. ACM 41(1), 181–204 (1994) 2. Bj¨ orklund, H., Boja´ nczyk, M.: Shuffle expressions and words with nested data. In: Kuˇcera, L., Kuˇcera, A. (eds.) MFCS 2007. LNCS, vol. 4708, pp. 750–761. Springer, Heidelberg (2007) 3. Bojanczyk, M., David, C., Muscholl, A., Schwentick, T., Segoufin, L.: Two-variable logic on data words. ACM Trans. Comput. Log. 12(4), 27 (2011) 4. Chambart, P., Schnoebelen, P.: The ordinal recursive complexity of lossy channel systems. In: LICS 2008, IEEE Symposium on Logic in Computer Science, IEEE Computer Society (2008) 5. Decker, N., Habermehl, P., Leucker, M., Thoma, D.: Ordered navigation on multiattributed data words. In: Baldan, P., Gorla, D. (eds.) CONCUR 2014. LNCS, vol. 8704, pp. 497–511. Springer, Heidelberg (2014) 6. Demri, S., Figueira, D., Praveen, M.: Reasoning about data repetitions with counter systems. In: LICS 2013, ACM/IEEE Symposium on Logic in Computer Science, IEEE Computer Society (2013) 7. Demri, S., Lazic, R.: LTL with the freeze quantifier and register automata. ACM Trans. Comput. Log. 10(3) (2009) 8. Demri, S., Lazic, R., Nowak, D.: On the freeze quantifier in constraint LTL: decidability and complexity. In: International Symposium on Temporal Representation and Reasoning (TIME 2005), IEEE Computer Society (2005)

284

N. Decker and D. Thoma

9. Figueira, D.: Alternating register automata on finite words and trees. Log. Meth. Comput. Sci. 8(1) (2012) 10. Fitting, M.: Modal logics between propositional and first-order. J. Log. Comput. 12(6), 1017–1026 (2002) 11. Goranko, V.: Hierarchies of modal and temporal logics with reference pointers. J. Logic Lang. Inf. 5(1), 1–24 (1996) 12. Haase, C., Schmitz, S., Schnoebelen, P.: The power of priority channel systems. Logical Meth. Comput. Sci. 10(4) (2014) 13. Henzinger, T.A.: Half-order modal logic: how to prove real-time properties. In: ACM Symposium on Principles of Distributed Computing, ACM (1990) 14. Kara, A., Schwentick, T., Zeume, T.: Temporal logics on words with multiple data values. In: FSTTCS 2010, IARCS Annual Conference on Foundations of Software Technology and Theoretical Computer Science, LIPIcs, vol. 8, Schloss Dagstuhl Leibniz-Zentrum f¨ ur Informatik (2010) 15. Lazi´c, R.S.: Safely freezing LTL. In: Arun-Kumar, S., Garg, N. (eds.) FSTTCS 2006. LNCS, vol. 4337, pp. 381–392. Springer, Heidelberg (2006) 16. Leroux, J., Schmitz, S.: Reachability in vector addition systems demystified. In: LICS 2015, ACM/IEEE Symposium on Logic in Computer Science, IEEE Computer Society (2015) 17. Lisitsa, A., Potapov, I.: Temporal logic with predicate lambda-abstraction. In: International Symposium on Temporal Representation and Reasoning (TIME 2005), IEEE Computer Society (2005) 18. Lomazova, I.A., Schnoebelen, P.: Some decidability results for nested petri nets. In: Bjorner, D., Broy, M., Zamulin, A.V. (eds.) PSI 1999. LNCS, vol. 1755, pp. 208–220. Springer, Heidelberg (2000) 19. Rosa-Velardo, F.: Ordinal recursive complexity of unordered data nets. Technical report TR-4-14, Departamento de Sistemas Inform´ aticos y Computaci´ on, Universidad Complutense de Madrid (2014) 20. Schmitz, S.: Complexity hierarchies beyond elementary. CoRR abs/1312.5686 (2013). http://arxiv.org/abs/1312.5686 21. Schnoebelen, P.: Revisiting Ackermann-Hardness for lossy counter machines and reset petri nets. In: Hlinˇen´ y, P., Kuˇcera, A. (eds.) MFCS 2010. LNCS, vol. 6281, pp. 616–628. Springer, Heidelberg (2010)

Regular Transformations of Data Words Through Origin Information Antoine Durand-Gasselin1 and Peter Habermehl2(B) 1

Aix Marseille Universit´e, CNRS, Centrale Marseille, LIF UMR 7279, Marseille, France [email protected] 2 IRIF, Univ. Paris Diderot & CNRS, Paris, France [email protected]

Abstract. We introduce a class of transformations of finite data words generalizing the well-known class of regular finite string transformations described by MSO-definable transductions of finite strings.These transformations map input words to output words whereas our transformations handle data words where each position has a letter from a finite alphabet and a data value. Each data value appearing in the output has as origin a data value in the input. As is the case for regular transformations we show that our class of transformations has equivalent characterizations in terms of deterministic two-way and streaming string transducers.

1

Introduction

The theory of transformations of strings (or words) over a finite alphabet has attracted a lot of interest recently. Courcelle [8] defined finite string transformations in a logical way using Monadic second-order definable graph transductions. Then, a breakthrough was achieved in [9] where it was shown that these transformations are equivalent to those definable by deterministic two-way finite transducers on finite words. In [1] deterministic streaming string transducers (SST) on finite words were introduced. This model is one-way but it is equipped with string variables allowing to store words. It is equivalent [1] to the deterministic two-way finite transducers and to MSO-definable transformations. Interestingly, the motivation behind SST was the more powerful SDST model [2]. SDST work on data words, i.e. words composed of couples of letters from a finite alphabet and an infinite data domain. However, they do not have the same nice theoretical properties as SST, for example they are not closed under composition because SDST have data variables allowing to store data values and compare data values with each other. Furthermore, there is no equivalent logical characterization. This work was supported in part by the VECOLIB project (ANR-14-CE28-0018) and by the PACS project (ANR-14-CE28-0002). A. Durand-Gasselin—Part of this work was done while this author was at Technical University Munich. c Springer-Verlag Berlin Heidelberg 2016  B. Jacobs and C. L¨ oding (Eds.): FOSSACS 2016, LNCS 9634, pp. 285–300, 2016. DOI: 10.1007/978-3-662-49630-5 17

286

A. Durand-Gasselin and P. Habermehl

In this paper, analogously to the case of string transformations of finite words, we obtain a class of transformations of finite data words which has an MSO characterization as well as equivalent characterizations in terms of deterministic two-way transducers and streaming transducers. To achieve this, we allow storing of data values in the transducers but not comparison. As an example we consider the transformation taking a finite input word over {#, a, b} starting and finishing with  a #, together with associated data values  from the integers, like #ab#abb# 1 24 5 671 4 and produces as output the word where (1) #’s are left unchanged, and between successive #’s (2) words w in {a, b}∗ are transformed into wR w where wR denotes the reverse image of w, and (3) the data value associated to each a is the value of the first # and the value   for b’s is the value of the second #. So, the output for the example word is #baab#bbaabb# 1 5115 5 445544 4 . It is clear how a deterministic two-way transducer with the ability of storing data values can realize this transformation: it stores the first data value (1) in   , then goes to the second #, stores the a data variable while outputting # 1 corresponding data value (5) in a second data variable, and goes back one by ba . Then, it turns around at the first #, goes again to the one while producing 51 ab second # producing 15 and restarts the same process. Now, to realize this transformation with a deterministic streaming string transducer one has to make with the fact that they can only go through the input word once from left to right. Nevertheless we will introduce a model which can realize the described transformation: in between two # s it stores the so-far read string and its reverse in a data word variable. As the data value of the p instead. For second # is not known in the beginning it uses a data parameter  baab . When reading the example, before the second #, the stored word will be p11p second #, it then replaces p everywhere by 5 and stores the result in another data word variable. The same repeats for the rest of the word until the end is reached and the output contained in a data word variable. The same transformation can also be described logically. To define transformations on data words, a natural choice would be to use transducers with origin information and their corresponding MSO transductions studied in [6]. Basically, their semantics also takes into account information about the origin of a letter in the output, i.e. the position in the input from which it originates. Obviously, this can be generalized to data values by defining the data value of each output position as the data value in the input position from where the output originated. This definition is however not expressive enough to handle our example, since an input position can only be the origin of a bounded number of output positions but the data values attached to (unboundedly many) a’s and b’s between two successive #’s come from the same two input positions. Therefore, in this paper, we first introduce a logical characterization of word transformations with generalized origin information. Our logical characterization is an extension of classical MSO transductions with an additional functional MSO defined relation that maps each element of the interpretation (symbols of the output word) to an element of the interpreted structure (symbols of the input word). This generalization naturally defines transformations of data words;

Regular Transformations of Data Words Through Origin Information

287

the data value at a given position of the output is the data value carried at the corresponding position in the input. This suffices to define the previously described example transformation. Interestingly, our class of transformations is captured by a natural model of deterministic two-way transducers extended with data variables whose values can neither be tested nor compared. By adding data word variables (as in streaming string transducers) containing data values and parameters, we then manage, while preserving determinism, to restrict that model to a one-way model. Data parameters are placeholders for data values which can be stored in data word variables and later replaced by concrete data values. We show that this one-way model can be captured by MSO to achieve the equivalence of all three models.

2 2.1

MSO Interpretations with MSO Origin Relation Words, Strings and Data Words

For S a set of symbols, we denote by S ∗ the set of finite words (i.e. the set of finite sequences of elements of S) over S. Given a word w, we can refer to its length (|w|), its first symbol (w[0]), the symbol at some position i < |w| in the word (w[i]), some of its subwords (e.g. w[i:j] with 0 ≤ i ≤ j < |w|, the subword between positions i and j) etc. In this paper, we only consider finite words. An alphabet (typically denoted Σ or Γ ) is a finite set of symbols. Furthermore, we use a (potentially infinite) set of data values called Δ. In the sequel, we use string to refer to a (finite) word over a finite alphabet and data word to refer to a word over the cartesian product of a finite alphabet and a set of data values. Since symbols of data words consist of a letter (from the finite alphabet) and a data value, we can naturally refer to the data value at some position in a data word, or the string corresponding to some data word. Conversely a string together with a mapping from its position to Δ forms a data word. A string w (over alphabet Σ) is naturally seen as a directed node-labeled graph (rather than considering edges only connecting two successive positions, we take the transitive closure: thus the graph is a finite linear order). The graph is then seen as an interpreted relational structure whose domain is the positions of w, with a binary edge predicate i. When reaching the end of the input word in a configuration (q, β), 

296

A. Durand-Gasselin and P. Habermehl

Fig. 1. The sstvp for the running example

a sstvp produces β(Ω(q)) if Ω(q) is defined. Then, naturally a sstvp S defines a transduction from words in Σ × Δ to words in Γ × Δ. The sstvp for our running example is given in Fig. 1. All data word variables are initialized with the empty word. By convention, a variable which is not explicitly updated is unchanged. We omit these updates for readability. Theorem 6. Equivalence of two sstvp is decidable. To prove this theorem we can generalize the proof of decidability of equivalence of SST [2], a reduction to reachability in a non-deterministic one-counter machine. Given two transducers we choose non-deterministically an input string, and one conflicting position in each of the two images (of the transducers): either they are labeled by different letters, or with attached data value originating from two distinct positions in the input word. We keep track in the counter of the difference between the number of produced symbols which will be part of each output before the corresponding conflicting position. Therefore, if the counter reaches 0 at the last letter of the input, the two transducers are different. We call SSTVP the class of all data word transductions definable by a sstvp. 4.2

From Two-Way to One-Way Transducers

Theorem 7. 2DFTV is included in SSTVP. Proof. (Sketch) We use ideas of [1] (based itself on Shepherdson’s translation from 2DFA to DFA [10]) where two-way transducers are translated into streaming string transducers. As they translate two-way transducers to copyless streaming string transducers they have to go through an intermediate model called heap-based transducers. Since we relax the copylessness restriction to 1-boundedness we can directly translate 2dftv to sstvp. Furthermore, we have to take care of the contents of data variables of the 2dftv. For that purpose we use data variables and data parameters of the sstvp. Since an sstvp does only one left-to-right pass on the input word, we cannot revisit any position. As we process a position we need to store all relevant information about that position possibly being later reprocessed by the two-way

Regular Transformations of Data Words Through Origin Information

297

transducer. The two-way transducer may process a position multiple times (each time in a different state) each time with a different valuation of data variables and producing some different word: for each state, we need to store in an appropriate data word variable the corresponding production, the valuation of data variables being abstracted with data parameters. Notice that not all these data word variables will be used in the output. Given a 2dftv A = (Σ, Γ, Δ, Q, q0 , F, V, μ, δ), over which we assume all accepting runs end on the last symbol, we define an sstvp B = (Σ, Γ, Δ, Q , X, V  , P, q0 , v0 , δ  , Ω) as follows: – Q = Q × [Q → (Q × 2V )] A state of the one-way transducer consists of a state of the two-way transducer and a partial mapping from states to a pair of a state and a set of variables. As a position i + 1 is processed, the state of B contains the following information: in which state A first reaches position i and for each state q of A what would be the state of A when it reaches for the first time position i + 1 had it started processing position i from state q: this part is the standard Shepherdson’s construction. The function is partial, as from position i from some states A might never reach position i + 1 (getting stuck). We remark that along the subrun from position i (in state q) to position i + 1, the A might store some data values in some data variables. The set of data variables denotes the set of data variables the two-way transducer has updated along that run. – X = xl ∪ {xq | q ∈ Q} At position i, variable xl will store the word produced by A until it first reaches position i. Variable xq will store the word produced from position i in state q until position i + 1 is first reached. – V  = V ∪ {vq | v ∈ V, q ∈ Q} At position i + 1, data variable v will contain the value of variable v of A as it first reaches position i + 1. Assume that B reaches position i in some state (q, f ) with f (q  ) = (q  , W ), and v ∈ W . Then variable vq will contain the last value stored in v when A processes from position i in state q  until it first reaches position i + 1. – P = {pv,q | v ∈ V, q ∈ Q} At position i, parameter pv,q will be present only in data word variable xq , representing that along the run of A the data value from data variable v at position i in state q was output before i + 1 was first reached. Such a symbol needs to be present in xq , but the data value is not yet known, hence it is abstracted by the data parameter pv,q . It is then easy to see how to define q0 and δ  so as to preserve these invariants. As B can not see , B must maintain the possible output in an extra variable, where it is supposed that the next symbol would be . We now detail an example (see Fig. 2) so as to give an intuition how δ  (q, σ) is built: we will specifically focus on the value of xq1 . We denote f the second component of q and we assume that f (q2 ) = (q3 , {v1 , v2 }), f (q4 ) = (q5 , {v2 , v3 }).

298

A. Durand-Gasselin and P. Habermehl

Fig. 2. An example to illustrate the transformation from A to B.

Furthermore, we assume that in A, δ(q1 , σ) = (q2 , −1, (γ, v2 )) and δ(q3 , σ) = (q4 , −1, (γ  , v2 )(γ  , v3 )) and finally that δ(q5 , σ) = (q6 , +1, (γ  , v2 )). Also reading σ in q1 and q3 assigns the current data value to v1 (i.e. μ(q1 , σ, v1 ) = μ(q3 , σ, v1 ) = curr), other data variables are not modified (i.e. μ(q1 , σ, vi ) = μ(q3 , σ, vi ) = vi ). By the aforementioned invariants, from state q1 , A will first reach the following position in state q6 (from the σ-labeled position in state q1 , it first goes left, reaches it again in state q3 , goes left again, arrives in state q5 and then moves to the right in state q6 ). If we abstract the data values, the content of the data word variable xq1 will thus be (γ, ?)xq2 (γ  , ?)(γ  , ?)xq4 (γ  , ?). Now we detail data attached to the produced letters, and the parameter assignments in the data word variables: γ will be given the data parameter p2,q1 . In xq2 : since a data value is assigned to v1 between q1 and q2 , p1,q1 should be substituted by that data value (which is curr) in xq2 . Other parameters in xq2 (which are all of the form pi,q2 ) are substituted by the corresponding pi,q1 . γ  will be given the data value v2,q2 and (because v3 has not been assigned a data value since q1 ) γ  will be assigned the data parameter p3,q1 . In xq4 : as a data value was assigned to v2 between q2 to q3 , parameter p2,q4 will be substituted by that value i.e. v2,q2 ; parameter p1,q4 will be substituted by curr and all other parameters (which are of the form pi,q4 ) will be assigned the corresponding data parameters pi,q1 . γ  should be assigned data value v2,q4 . Therefore by reading a σ in B, we reach a state whose second component maps q1 to (q6 , {v1 , v2 , v3 }), v1,q1 ← curr, v2,q1 ← v2,q4 , v3,q1 ← v3,q4 . 4.3

From One-Way Transducers to MSO

In order to conclude that the three models of data word transformations are equivalent, it remains to show that our MSO transductions with MSO origin information capture all transformations defined by the one-way model. Theorem 8. SSTVP is included in MSOT+O.

Regular Transformations of Data Words Through Origin Information

299

The proof is very similar to that of encoding finite state automata in MSO. Usually to show that MSO captures string transformations defined by a one-way model one defines an output graph with Γ -labeled edges and -edges. We directly give a proof that builds a (string) graph whose nodes are Γ -labeled. Given an sstvp S we fix the set of copies C as the set of occurrences of symbols of Γ in the variable update function. Since S is deterministic, we will write an MSO sentence ϕ that characterizes a run of a word in S. This formula will be of the form ∃X1 , . . . Xn ψ, such that given a word w (in the domain of the transformation), there exists a unique assignment of the Xi such that ψ holds. These second order variables consist of: – Xq for q ∈ Q: position i ∈ Xq iff processing position i yielded state q. – Xr for every word variable r: position i ∈ Xr iff the content of variable r will flow in the output – Xr1 ,r2 for every pair of distinct word variables r1 , r2 : position i ∈ Xr1 ,r2 iff the content of variable r1 will flow in the output before the content of the variable r2 that will also flow in the output. Our sequential machine model allows easily to write such a formula ψ. With the formula ψ, we can write formulas ϕindom , (ϕcdom )c∈C , (ϕcγ )c∈C , and (ϕc,d < )c,d∈C . We remark that second order variables Xr1 ,r2 have a unique valid assignment because of the (semantic) copylessness of sstvp. These variables are typically used to define ϕc,d < . To hint how to build formula ϕcorig (x, y) we state the following simple lemma about runs of sstvps. Lemma 2. Given an sstvp S, an input word w and position x that produces a symbol γ ∈ Γ that will be part of the output. – Either γ is produced with a data variable (namely v): In this case, there exists a unique position y ≤ x where the data value curr was stored in some data variable and that data variable flows to data variable v at position x. – or γ is produced with a data parameter (namely p): In this case, there exists a unique position z such that the data parameter attached to γ is some pm at position z and that pm is assigned a variable vm (or curr) at position z. There exists a unique position y ≤ q such that at position y the data value curr was put in some data variable, which flows to a data variable vm at position z. The notion of “flow” is easily expressed with ψ and second order existential quantification. The copyless semantics of sstvps ensures that to each (output) symbols, exactly one data value (or equivalently a unique position from the input word) is assigned to. This allows to build MSO formulas ϕcorig that have the desired functional property.

300

5

A. Durand-Gasselin and P. Habermehl

Conclusion

Finite string transformation have been generalized to infinite string transformations [5] and tree transformations [3,4]. It would be interesting to extend our results to these settings by adding data values and defining transformations via origin information. Furthermore, it would be interesting to study the pre-post condition checking problem along the lines of [2], i.e. the problem to check that given a transducer is it the case that each input satisfying a pre-condition defined via some automata-model is transformed into an output satisfying a similarly defined post-condition.

References ˇ 1. Alur, R., Cern´ y, P.: Expressiveness of streaming string transducers. In: FSTTCS, vol. 8, pp. 1–12 (2010) ˇ 2. Alur, R., Cern´ y, P.: Streaming transducers for algorithmic verification of singlepass list-processing programs. In: POPL, pp. 599–610 (2011) 3. Alur, R., D’Antoni, L.: Streaming tree transducers. In: Czumaj, A., Mehlhorn, K., Pitts, A., Wattenhofer, R. (eds.) ICALP 2012, Part II. LNCS, vol. 7392, pp. 42–53. Springer, Heidelberg (2012) 4. Alur, R., Durand-Gasselin, A., Trivedi, A.: From monadic second-order definable string transformations to transducers. In: 28th Annual ACM/IEEE Symposium on Logic in Computer Science, LICS, pp. 458–467. IEEE Computer Society (2013) 5. Alur, R., Filiot, E., Trivedi, A.: Regular transformations of infinite strings. In: Proceedings of the 27th Annual IEEE Symposium on Logic in Computer Science, LICS, pp. 65–74. IEEE Computer Society (2012) 6. Boja´ nczyk, M.: Transducers with origin information. In: Esparza, J., Fraigniaud, P., Husfeldt, T., Koutsoupias, E. (eds.) ICALP 2014, Part II. LNCS, vol. 8573, pp. 26–37. Springer, Heidelberg (2014) 7. Chytil, M., J´ akl, V.: Serial composition of 2-way finite-state transducers and simple programs on strings. In: Salomaa, A., Steinby, M. (eds.) Automata, Languages and Programming. LNCS, vol. 52, pp. 135–147. Springer, London (1977) 8. Courcelle, B.: Monadic second-order definable graph transductions: a survey. Theoret. Comput. Sci. 126(1), 53–75 (1994) 9. Engelfriet, J., Hoogeboom, H.J.: MSO definable string transductions and two-way finite-state transducers. ACM Trans. Comput. Log. 2, 216–254 (2001) 10. Shepherdson, J.C.: The reduction of two-way automata to one-way automata. IBM J. Res. Dev. 3(2), 198–200 (1959) 11. Thomas, W.: Ehrenfeucht games, the composition method, and the monadic theory of ordinal words. In: Mycielski, J., Rozenberg, G., Salomaa, A. (eds.) Structures in Logic and Computer Science. LNCS, vol. 1261, pp. 118–143. Springer, Heidelberg (1997)

Probabilistic and Timed Systems

Trace Refinement in Labelled Markov Decision Processes Nathana¨el Fijalkow(B) , Stefan Kiefer, and Mahsa Shirmohammadi University of Oxford, Oxford, UK {nathanael.fijalkow,stefan.kiefer,mahsa.shirmohammadi}@cs.ox.ac.uk

Abstract. Given two labelled Markov decision processes (MDPs), the trace-refinement problem asks whether for all strategies of the first MDP there exists a strategy of the second MDP such that the induced labelled Markov chains are trace-equivalent. We show that this problem is decidable in polynomial time if the second MDP is a Markov chain. The algorithm is based on new results on a particular notion of bisimulation between distributions over the states. However, we show that the general trace-refinement problem is undecidable, even if the first MDP is a Markov chain. Decidability of those problems was stated as open in 2008. We further study the decidability and complexity of the trace-refinement problem provided that the strategies are restricted to be memoryless.

1

Introduction

We consider labelled Markov chains (MCs) whose transitions are labelled with symbols from an alphabet L. Upon taking a transition, the MC emits the associated label. In this way, an MC defines a trace-probability function Tr : L∗ → [0, 1] which assigns to each finite trace w ∈ L∗ the probability that the MC emits w during its first |w| transitions. Consider the MC depicted in Fig. 1 with initial state p0 . For example, see that if in state p0 , with probability 14 , a transition to state pc is taken and c is emitted. We have, e.g., Tr (abc) = 14 · 14 · 14 . Two MCs over the same alphabet L are called equivalent if their trace-probability functions are equal. The study of such MCs and their equivalence has a long history, going back to Sch¨ utzenberger [19] and Paz [16]. Sch¨ utzenberger and Paz studied weighted and probabilistic automata, respectively. Those models generalize labelled MCs, but the respective equivalence problems are essentially the same. It can be extracted from [19] that equivalence is decidable in polynomial time, using a technique based on linear algebra. Variants of this technique were developed, see e.g. [8, 20]. Tzeng [21] considered the path-equivalence problem for nondeterministic automata which asks, given nondeterministic automata A and B, whether each word has the same number of accepting paths in A as in B. He gives an NC algorithm1 for deciding path equivalence which can be straightforwardly adapted to yield an NC algorithm for equivalence of MCs. 1

The complexity class NC is the subclass of P containing those problems that can be solved in polylogarithmic parallel time (see e.g. [10]).

c Her Majesty the Queen in Right of the United Kingdom 2016  B. Jacobs and C. L¨ oding (Eds.): FOSSACS 2016, LNCS 9634, pp. 303–318, 2016. DOI: 10.1007/978-3-662-49630-5 18

304

N. Fijalkow et al.

Fig. 1. An MC with its trace-probability function. This MC, denoted by C(A), is also used in the reduction from universality of probabilistic automata to MC  MDP.

Fig. 2. An MDP where the choice of controller is relevant only in q1 . Two available moves m1 , m2 are shown with small black circles.

More recently, the efficient decidability of the equivalence problem was exploited, both theoretically and practically, for the verification of probabilistic systems, see e.g. [12–15, 17]. In those works, equivalence naturally expresses properties such as obliviousness and anonymity, which are difficult to formalize in temporal logic. The inclusion problem for two probabilistic automata asks whether for each word the acceptance probability in the first automaton is less than or equal to the acceptance probability in the second automaton. Despite its semblance to the equivalence problem, the inclusion problem is undecidable [6], even for automata of fixed dimension [3]. This is unfortunate, especially because deciding language inclusion is often at the heart of verification algorithms. We study another “inclusion-like” generalization of the equivalence problem: trace refinement in labelled Markov decision processes (MDPs). MDPs extend MCs by nondeterminism; in each state, a controller chooses, possibly randomly and possibly depending on the history, one out of finitely many moves 2 . A move determines a probability distribution over the emitted label and the successor state. In this way, an MDP and a strategy of the controller induce an MC. The trace-refinement problem asks, given two MDPs D and E, whether for all strategies for D there is a strategy for E such that the induced MCs are equivalent. Consider the MDP depicted in Fig. 2 where in state q1 there are two available moves; one move generates the label c with probability 1, the other move generates d with probability 1. A strategy of the controller that chooses the last generated label in the state q1 , either c or d, with probability 1, induces the same trace-probability function as the MC shown in Fig. 1; the MDP thus refines that MC. The described strategy needs one bit of memory to keep track of the last generated label. It was shown in [8] that the strategy for E may require infinite memory, even if D is an MC. The decidability of trace refinement was posed as an open problem, both in the introduction and in the conclusion of [8]. The authors of [8] also ask about the decidability of subcases, where D or E are restricted to be MCs. In this paper we answer all those questions. We show 2

As in [8] we speak of moves rather than of actions, to avoid possible confusion with the label alphabet L.

Trace Refinement in Labelled Markov Decision Processes

305

that trace refinement is undecidable, even if D is an MC. In contrast, we show that trace refinement is decidable efficiently (in NC, hence in P), if E is an MC. Moreover, we prove that the trace-refinement problem becomes decidable if one imposes suitable restrictions on the strategies for D and E, respectively. More specifically, we consider memoryless (i.e., no dependence on the history) and pure memoryless (i.e., no randomization and no dependence on the history) strategies, establishing various complexity results between NP and PSPACE. To obtain the aforementioned NC result, we demonstrate a link between trace refinement and a particular notion of bisimulation between two MDPs that was studied in [11]. This variant of bisimulation is not defined between two states as in the usual notion, but between two distributions on states. An exponentialtime algorithm that decides (this notion of) bisimulation was provided in [11]. We sharpen this result by exhibiting a coNP algorithm that decides bisimulation between two MDPs, and an NC algorithm for the case where one of the MDPs is an MC. For that we refine the arguments devised in [11]. The model considered in [11] is more general than ours in that they also consider continuous state spaces, but more restricted than ours in that the label is determined by the move. The full version of the paper is available online [1].

2

Preliminaries

A trace over a finite set L of labels is a finite sequence w = a1 · · · an of labels where the length of the trace is |w| = n. The empty trace  has length zero. For n ≥ 0, let Ln be the set of all traces with length n; we denote by L∗ the set of all (finite) traces over L.  For a function d : S → [0, 1] over a finite set S, define the norm d := s∈S d(s). The support of d is the set Supp(d) = {s ∈ S | d(s) > 0}. The function d is a probability subdistribution over S if d ≤ 1; it is a probability distribution if d = 1. We denote by subDist(S) (resp. Dist(S)) the set of all probability subdistributions (resp. distributions) over S. Given s ∈ S, the Dirac distribution on s assigns probability 1 to s; we denote it by ds . For a non-empty subset T ⊆ S, the uniform distribution over T assigns probability |T1 | to every element in T . 2.1

Labelled Markov Decision Processes

A labelled Markov decision process (MDP) D = Q, μ0 , L, δ consists of a finite set Q of states, an initial distribution μ0 ∈ Dist(Q), a finite set L of labels, and a finite probabilistic transition relation δ ⊆ Q × Dist(L × Q) where states are in relation with distributions over pairs of labels and successors. We assume that for each state q ∈ Q there exists some distribution d ∈ Dist(L × Q) where q, d ∈ δ. The set of  moves in q is moves(q) = {d ∈ Dist(L × Q) | q, d ∈ δ}; denote by moves = q∈Q moves(q) the set of all moves.

306

N. Fijalkow et al.

For the complexity results, we assume that probabilities of transitions are rational and given as fractions of integers represented in binary. We describe the behaviour of an MDP as a trace generator running in steps. The MDP starts in the first step in state q with probability μ0 (q). In each step, if the MDP is in state q the controller chooses m ∈ moves(q); then, with probability m(a, q  ), the label a is generated and the next step starts in the successor state q  . Given q ∈ Q, denote by post(q) the set {(a, q  ) ∈ Supp(m) | m ∈ moves(q)}. A path in D is a sequence ρ = q0 a1 q1 . . . an qn such that (ai+1 , qi+1 ) ∈ post(qi ) for all 0 ≤ i < n. The path ρ has the last state last(ρ) = qn ; and the generated trace after ρ is a1 a2 · · · an , denoted by trace(ρ). We denote by Paths(D) the set of all paths in D, and by Paths(w) = {ρ ∈ Paths(D) | trace(ρ) = w} the set of all path generating w. Strategies. A randomized strategy (or simply a strategy) for an MDP D is a function α : Paths(D) → Dist(moves) that, given a finite path ρ, returns a probability distribution α(ρ) ∈ Dist(moves(last(ρ))) over the set of moves in last(ρ), used to generate a label a and select a successor state q  with probability   m∈moves(q) α(ρ)(m) · m(a, q ) where q = last(ρ). A strategy α is pure if for all ρ ∈ Paths(D), we have α(ρ)(m) = 1 for some m ∈ moves; we thus view pure strategies as functions α : Paths(D) → moves. A strategy α is memoryless if α(ρ) = α(ρ ) for all paths ρ, ρ with last(ρ) = last(ρ ); we thus view memoryless strategies as functions α : Q → Dist(moves). A strategy α is trace-based if α(ρ) = α(ρ ) for all ρ, ρ where trace(ρ) = trace(ρ ) and last(ρ) = last(ρ ); we view trace-based strategies as functions α : L∗ × Q → Dist(moves). Trace-Probability Function. For an MDP D and a strategy α, the probability of a single path is inductively defined by PrD,α (q) = μ0 (q) and  PrD,α (ρaq) = PrD,α (ρ) · α(ρ)(m) · m(a, q). m∈moves(last(ρ))

The trace-probability function Tr D,α : L∗ → [0, 1] is, given a trace w, defined by Tr D,α (w) =



PrD,α (ρ).

ρ∈Paths(w)

We may drop the subscript D or α from Tr D,α if it is understood. We denote by subDis D,α (w) ∈ subDist(Q), the subdistribution after generating a traces w, that is  PrD,α (ρ). subDis D,α (w)(q) = ρ∈Paths(w):last(ρ)=q

We have: Tr D,α (w) = subDis D,α (w) A version of the following lemma was proved in [8, Lemma 1]:

(1)

Trace Refinement in Labelled Markov Decision Processes

307

Lemma 1. Let D be an MDP and α be a strategy. There exists a trace-based strategy β such that Tr α = Tr β . Here, by Tr α = Tr β we mean Tr α (w) = Tr β (w) for all traces w ∈ L∗ . Labeled Markov Chains. A finite-state labeled Markov chain (MC for short) is an MDP where only a single move is available in each state, and thus controller’s choice plays no role. An MC C = Q, μ0 , L, δ is an MDP where δ : Q → Dist(L × Q) is a probabilistic transition function. Since MCs are MDPs, we analogously define paths, and the probability of a single path inductively as follows: PrC (q) = μ0 (q) and PrC (ρaq) = PrC (ρ) · δ(q  )(a, q) where q  = last(ρ). The notations subDis C (w) and Tr C are defined analogously. 2.2

Trace Refinement

Given two MDPs D and E with the same set L of labels, we say that E refines D, denoted by D E, if for all strategies α for D there exists some strategy β for E such that Tr D = Tr E . We are interested in the problem MDP MDP, which asks, for two given MDPs D and E, whether D E. The decidability of this problem was posed as an open question in [8]. We show in Theorem 2 that the problem MDP MDP is undecidable. We consider various subproblems of MDP MDP, which asks whether D E holds. Specifically, we speak of the problem – MDP MC when E is restricted to be an MC; – MC MDP when D is restricted to be an MC; – MC MC when both D and E are restricted to be MCs. We show in Theorem 2 that even the problem MC MDP is undecidable. Hence we consider further subproblems. Specifically, we denote by MC MDPm the problem where the MDP is restricted to use only memoryless strategies, and by MC MDPpm the problem where the MDP is restricted to use only pure memoryless strategies. When both MDPs D and E are restricted to use only pure memoryless strategies, the trace-refinement problem is denoted by MDPpm MDPpm . The problem MC MC equals the trace-equivalence problem for MCs: given two MCs C1 , C2 we have C1 C2 if and only if Tr C1 = Tr C2 if and only if C2 C1 . This problem is known to be in NC [21], hence in P.

3

Undecidability Results

In this section we show: Theorem 2. The problem MC MDP is undecidable. Hence a fortiori, MDP MDP is undecidable.

308

N. Fijalkow et al.

Fig. 3. Sketch of the construction of the MDP D from the probabilistic automaton A, for the undecidability result of MC  MDP. Here, p is an accepting state whereas q is not. To read the picture, note that in p there is a transition to the state p1 with probability x and label a: δ(p, a)(p1 ) = x.

Proof. To show that the problem MC MDP is undecidable, we establish a reduction from the universality problem for probabilistic automata. A probabilistic automaton is a tuple A = Q, μ0 , L, δ, F consisting of a finite set Q of states, an initial distribution μ0 ∈ Dist(Q), a finite set L of letters, a transition function δ : Q × L → Dist(Q) assigning to every state and letter a distribution over states, and a set F of final states. For a word w ∈ L∗ we write dis A (w) ∈ Dist(Q) for the distribution such that, for all q ∈ Q, we have that that, after inputting w, the automaton A is in dis A (w)(q) is the probability  state q. We write PrA (w) = q∈F dis A (w)(q) to denote the probability that A accepts w. The universality problem asks, given a probabilistic automaton A, whether PrA (w) ≥ 12 holds for all words w. This problem is known to be undecidable [16]. Let A = Q, μ0 , L, δ, F be a probabilistic automaton; without loss of generality we assume that L = {a, b}. We construct an MDP D such that A is universal if and only if C D where C is the MC shown in Fig. 1. The MDP D is constructed from A as follows; see Fig. 3. Its set of states is Q ∪ {qc , qd }, and its initial distribution is μ0 . (Here and in the following we identify subdistributions μ ∈ subDist(Q) and μ ∈ subDist(Q ∪ {qc , qd }) if μ(qc ) = μ(qd ) = 0.) We describe the transitions of D using the transition function δ of A. Consider a state q ∈ Q: – If q ∈ F , there are two available moves mc , md ; both emit a with probability 1 4 and simulate the probabilistic automaton A reading the letter a, or emit b with probability 14 and simulate the probabilistic automaton A reading the letter b. With the remaining probability of 12 , mc emits c and leads to qc and md emits d and leads to qd . Formally, mc (c, qc ) = 12 , md (d, qd ) = 12 and mc (e, q  ) = md (e, q  ) = 14 δ(q, e)(q  ) where q  ∈ Q and e ∈ {a, b}.

Trace Refinement in Labelled Markov Decision Processes

309

– If q ∈ / F , there is a single available move m such that m(d, qd ) = 12 and m(e, q  ) = 14 δ(q, e)(q  ) where q  ∈ Q and e ∈ {a, b}. – The only move from qc is the Dirac distribution on (c, qc ); likewise the only move from qd is the Dirac distribution on (d, qd ). This MDP D “is almost” an MC, in the sense that a strategy α does not influence its behaviour until eventually a transition to qc or qd is taken. Indeed, for all α 1 dis A (w). In particular, it and for all w ∈ {a, b}∗ we have subDis D,α (w) = 4|w| 1 1 . Further, if α is follows Tr D,α (w) = subDis D,α (w) = 4|w| dis A (w) = 4|w| trace-based we have: Tr D,α (wc) = subDis D,α (wc) = subDis D,α (wc)(qc )  1 = subDis D,α (w)(q) · α(w, q)(mc ) · 2

by (1) structure of D structure of D

q∈F

=

1 

4|w|

dis A (w)(q) · α(w, q)(mc ) ·

q∈F

1 2

(2)

as argued above

We show that A is universal if and only if C D. Let A be universal. Define a trace-based strategy α with α(w, q)(mc ) = 2PrA1 (w) for all w ∈ {a, b}∗ and q ∈ F . Note that α(w, q)(mc ) is a probability as PrA (w) ≥ 12 . Let w ∈ {a, b}∗ . We have: Tr D,α (w) =

1

as argued above

4|w|

= Tr C (w)

Figure 1

Further we have: Tr D,α (wc) = = =

1  4|w|

q∈F

1  4|w|

dis A (w)(q) · α(w, q)(mc ) ·

q∈F

1 4|w|+1

= Tr C (wc)

dis A (w)(q) ·

1 1 · PrA (w) 4

1 2

by (2) definition of α PrA (w) =



dis A (w)(q)

q∈F

Figure 1

It follows from the definitions of D and C that for all k ≥ 1, we have Tr D,α (wck ) = Tr D,α (wc) = Tr C (wc) = Tr C (wck ). We have  e∈{a,b,c,d} Tr D,α (we) = Tr D,α (w) = Tr C (w) = e∈{a,b,c,d} Tr C (we). Since for e ∈ {a, b, c} we also proved that Tr D,α (we) = Tr C (we) it follows that Tr D,α (wd) = Tr C (wd). Hence, as above, Tr D,α (wdk ) = Tr C (wdk ) for all k ≥ 1. Finally, if w ∈ / (a + b)∗ · (c∗ + d∗ ) then Tr D,α (w) = 0 = Tr C (w).

310

N. Fijalkow et al.

For the converse, assume that A is not universal. Then there is w ∈ {a, b}∗ with PrA (w) < 12 . Let α be a trace-based strategy. Then we have: Tr D,α (wc) = ≤ =

1  4|w| 1 4|w| 1 4|w|

dis A (w)(q) · α(w, q)(mc ) ·

q∈F

·

1

4|w|

by (2)

1  · dis A (w)(q) 2

α(w, q)(mc ) ≤ 1

1 · PrA (w) 2

PrA (w) =

q∈F

·

1 1 · 2 2 = Tr C (wc)
0, where playing σ ensures a probability ≥ 1 −  of reaching a given target state, but no strategy ensures probability 1. In this case, one says that the reachability property holds limit-surely, but not almost-surely (i.e., unlike in finite-state MDPs, almost-surely and limit-surely do not coincide in infinite-state MDPs). We show that even for P-VASS-MDPs, all sure/almost-sure/limit-sure reachability/B¨ uchi problems are still undecidable. However, in the deadlock-free subclass of P-VASS-MDPs, the sure reachability/B¨ uchi problems become decidable (while the other problems remain undecidable). In contrast, for 1-VASS-MDPs, the sure/almost-sure/limit-sure reachability problem and the sure/almost-sure B¨ uchi problem are decidable. Our decidability results rely on two different techniques. For the sure and almost sure problems, we prove that we can reduce them to the model-checking problem over VASS of a restricted fragment of the modal μ-calculus that has been proved to be decidable in [4]. For the limit-sure reachability problem in 1VASS-MDP, we use an algorithm which at each iteration reduces the dimension of the considered VASS while preserving the limit-sure reachability properties. Although we do not consider the class of qualitative objectives referring to the probability of (repeated) reachability being strictly greater than 0, we observe that reachability on VASS-MDPs in such a setting is equivalent to reachability on standard VASS (though this correspondence does not hold for repeated reachability).

Qualitative Analysis of VASS-Induced MDPs

321

Outline. In Sect. 2 we define basic notations and how VASS induce MDPs. In Sects. 3 and 4 we consider verification problems for P-VASS-MDP and 1-VASSMDP, respectively. In Sect. 5 we summarize the decidability results (Table 1) and outline future work. Omitted proofs can be found in [2].

2

Models and Verification Problems

Let N (resp. Z) denote the set of nonnegative integers (resp. integers). For two integers i, j such that i ≤ j we use [i..j] to represent the set {k ∈ Z | i ≤ k ≤ j}. Given a set X and n ∈ N \ {0}, X n is the set of n-dimensional vectors with values in X. We use 0 to denote the vector such that 0(i) = 0 for all i ∈ [1..n]. The classical order on Zn is denoted ≤ and is defined by v ≤ w if and only if v(i) ≤ w(i) for all i ∈ [1..n]. We also define the operation + over n-dimensional vectors of integers in the classical way (i.e., for v, v ∈ Zn , v + v is defined by (v + v )(i) = v(i) + v (i) for all i ∈ [1..n]). Given a set S, we use S ∗ (respectively S ω ) to denote the set of finite (respectively infinite) sequences of elements of S. We now recall the notion of well-quasi-ordering (which we abbreviate as wqo). A quasi-order (A, ) is a wqo if for every infinite sequence of elements a1 , a2 , . . . in A, there exist two indices i < j such that ai  aj . For n > 0, (Nn , ≤) is a wqo. Given a set A with an ordering  and a subset B ⊆ A, the set B is said to be upward closed in A if a1 ∈ B, a2 ∈ A and a1  a2 implies a2 ∈ B. 2.1

Markov Decision Processes

A probability  distribution on a countable set X is a function f : X → [0, 1] such that x∈X f (x) = 1. We use D(X) to denote the set of all probability distributions on X. We first recall the definition of Markov decision processes. Definition 1 (MDPs). A Markov decision process (MDP) M is a tuple

C, C1 , CP , A, →, p where: C is a countable set of configurations partitioned into C1 and CP (that is C = C1 ∪ CP and C1 ∩ CP = ∅); A is a set of actions; →⊆ C × A × C is a transition relation; p : CP → D(C) is a partial function which assigns to some configurations in CP probability distributions on C such a → c for some a ∈ A. that p(c)(c ) > 0 if and only if c − Note that our definition is equivalent as seeing MDPs as games played between a nondeterministic player (Player 1) and a probabilistic player (Player P). The set C1 contains the nondeterministic configurations (or configurations of Player 1) and the set CP contains the probabilistic configurations (or configurations of Player P). Given two configurations c, c in C, we write c → c a whenever there exists a ∈ A such that c − → c . We will say that a configuration c ∈ C is a deadlock if there does not exist c ∈ C such that c → c . We use C1df (resp. CPdf ), to denote the configurations of Player 1 (resp. of Player P) which are not a deadlock (df stands here for deadlock free).

322

P.A. Abdulla et al.

A play of the MDP M = C, C1 , CP , A, →, p is either an infinite sequence of ak−1 a0 a1 a0 a1 the form c0 −→ c1 −→ c2 · · · or a finite sequence c0 −→ c1 −→ c2 · · · −−−→ ck . We call the first kind of play an infinite play, and the second one a finite play. A play is said to be maximal whenever it is infinite or it ends in a deadlock configuration. These latter plays are called deadlocked plays. We use Ω to denote the set of ak−1 a0 a1 c1 −→ c2 · · · −−−→ ck , let ck = last(ρ). maximal plays. For a finite play ρ = c0 −→ df We use Ω1 to denote the set of finite plays ρ such that last(ρ) ∈ C1df . A strategy for Player 1 is a function σ : Ω1df → C such that, for all ρ ∈ Ω1df and c ∈ C, if σ(ρ) = c then last(ρ) → c. Intuitively, given a finite play ρ, which represents the history of the game so far, the strategy represents the choice of Player 1 among the different possible successor configurations from last(ρ). We use Σ to denote the set of all strategies for Player 1. Given a strategy σ ∈ Σ, a0 a1 c1 −→ c2 · · · respects σ if for every k ∈ N, we have that if an infinite play c0 −→ a0 a1 c1 −→ c2 · · · ck ) and if ck ∈ CP then p(ck )(ck+1 ) > ck ∈ C1 then ck+1 = σ(c0 −→ 0. We define finite plays that respect σ similarly. Let Plays(M, c, σ) ⊆ Ω be the set of all maximal plays of M that start from c and that respect σ. Note that once a starting configuration c0 ∈ C and a strategy σ have been chosen, the MDP is reduced to an ordinary stochastic process. We define an event A ⊆ Ω as a measurable set of plays and we use P(M, c, σ, A) to denote the probability of event A starting from c ∈ C under strategy σ. The notation P+ (M, c, A) will be used to represent the maximal probability of event A starting from c which is defined as P+ (M, c, A) = supσ∈Σ P(M, c, σ, A). 2.2

VASS-MDPs

Probabilistic Vector Addition Systems with States have been studied, e.g., in [3]. Here we extend this model with non-deterministic choices made by a controller. We call this new model VASS-MDPs. We first recall the definition of Vector Addition Systems with States. Definition 2 (VASS). For n > 0, an n-dimensional Vector Addition System with States (VASS) is a tuple S = Q, T where Q is a finite set of control states and T ⊆ Q × Zn × Q is the transition relation labelled with vectors of integers. In the sequel, we will not always make precise the dimension of the considered VASS. Configurations of a VASS are pairs q, v ∈ Q×Nn . Given a configuration

q, v and a transition t = q, z, q  in T , we will say that t is enabled at q  , v , if q = q  and v+z ≥ 0. Let then En(q, v) be the set {t ∈ T | t is enabled at q, v) }. In case the transition t = q, z, q  is enabled at q, v , we define t(q, v) = q  , v where v = v+z. An n-dimensional VASS S induces a labelled transition system

C, T, → where C = Q × Nn is the set of configurations and the transition t relation →⊆ C × T × C is defined as follows: q, v − → q  , v iff q  , v = t(q, v). VASS are sometimes seen as programs manipulating integer variables, a.k.a. counters. When a transition of a VASS changes the i-th value of a vector v, we will sometimes say that it modifies the value of the i-th counter. We show now in which manner we add probability distributions to VASS.

Qualitative Analysis of VASS-Induced MDPs

323

Definition 3 (VASS-MDP). A VASS-MDP is a tuple S = Q, Q1 , QP , T, τ where Q, T is a VASS for which the set of control states Q is partitioned into Q1 and QP , and τ : T → N\{0} is a partial function assigning to each transition a weight which is a positive natural number. Nondeterministic (resp. probabilistic) choices are made from control states in Q1 (resp. QP ). The subset of transitions from control states of Q1 (resp. control states of QP ) is denoted by T1 (resp. TP ). Hence T = T1 ∪ TP with T1 ⊆ Q1 ×Zn ×Q and TP ⊆ QP ×Zn ×Q. A VASS-MDP S = Q, Q1 , QP , T, τ induces an MDP MS = C, C1 , CP , T, →, p where: C, T, → is the labelled transition system associated with the VASS Q, T ; C1 = Q1 × Nn and CP = QP × Nn ;   and for all c ∈ CPdf and c ∈ C, if c → c , the  probability of going from c to c  is defined by p(c)(c ) = ( {t|t(c)=c } τ (t))/( t∈En(c) τ (t)), whereas if c → c , we have p(c)(c ) = 0. Note that the MDP MS is well-defined: when defining p(c)(c )  in the case c → c , there exists at least one transition in En(c) and consequently the sum t∈En(c) τ (t) is never equal to 0. Also, we could have restricted the weights to be assigned only to transitions leaving from a control state in QP since we do not take into account the weights assigned to the other transitions. A VASS-MDP is deadlock free if its underlying VASS is deadlock free. Finally, as in [18] or [4], we will see that to gain decidability it is useful to restrict the power of the nondeterministic player or of the probabilistic player by restricting their ability to modify the counters’ values and hence letting them only choose a control location. This leads to the two following definitions: a PVASS-MDP is a VASS-MDP Q, Q1 , QP , T, τ such that for all q, z, q  ∈ T1 , we have z = 0 and a 1-VASS-MDP is a VASS-MDP Q, Q1 , QP , T, τ such that for all q, z, q  ∈ TP , we have z = 0. In other words, in a P-VASS-MDP, Player 1 cannot change the counter values when taking a transition and, in a 1-VASS-MDP, it is Player P which cannot perform such an action. 2.3

Verification Problems for VASS-MDPs

We consider qualitative verification problems for VASS-MDPs, taking as objectives control-state reachability and repeated reachability. To simplify the presentation, we consider a single target control-state qF ∈ Q. However, our positive decidability results easily carry over to sets of target control-states (while the negative ones trivially do). Note however, that asking to reach a fixed target configuration like qF , 0 is a very different problem (cf. [3]). Let S = Q, Q1 , QP , T, τ be a VASS-MDP and MS its associated MDP. Given a control state qF ∈ Q, we denote by ♦qF  the set of infinite plays c0 · c1 · · · · and deadlocked plays c0 · · · · · cl of MS for which there exists an index k ∈ N such that ck = qF , v for some v ∈ Nn . Similarly, ♦qF  characterizes the set of infinite plays c0 · c1 · · · · of MS for which the set {i ∈ N | ci =

qF , v for some v ∈ Nn } is infinite. Since MS is an MDP with a countable number of configurations, we know that the sets of plays ♦qF  and ♦qF  are measurable (for more details see for instance [5]), and are hence events for

324

P.A. Abdulla et al.

Fig. 1. Two 1-dimensional VASS-MDPs. The circles (resp. squares) are the control states of Player 1 (resp. Player P). All transitions have the same weight 1. From q0 , 0, the state qF is reached almost-surely, but not surely, due to the possible run with an infinite loop at q0 (which has probability zero). From q1 , 0, the state qF can be reached limit-surely (by a family of strategies that repeats the loop at q1 more and more often), but not almost-surely (or surely), since every strategy has a chance of getting stuck at state q2 with counter value zero.

MS . Given an initial configuration c0 ∈ Q × Nn and a control-state qF ∈ Q, we consider the following questions: 1. The sure reachability problem: Does there exist a strategy σ ∈ Σ such that Plays(MS , c0 , σ) ⊆ ♦qF ? 2. The almost-sure reachability problem: Does there exist a strategy σ ∈ Σ such that P(MS , c0 , σ, ♦qF ) = 1? 3. The limit-sure reachability problem: Does P+ (MS , c0 , ♦qF ) = 1? 4. The sure repeated reachability problem: Does there exist a strategy σ ∈ Σ such that Plays(MS , c0 , σ) ⊆ ♦qF ? 5. The almost-sure repeated reachability problem: Does there exist a strategy σ ∈ Σ such that P(MS , c0 , σ, ♦qF ) = 1? 6. The limit-sure repeated reachability problem: Does P+ (MS , c0 , ♦qF ) = 1? Note that sure reachability implies almost-sure reachability, which itself implies limit-sure reachability, but not vice-versa, as shown by the counterexamples in Fig. 1 (see also [7]). The same holds for repeated reachability. For the sure problems, probabilities are not taken into account, and thus these problems can be interpreted as the answer to a two player reachability game played on the transition system of S. Such games have been studied for instance in [1, 4, 18]. Finally, VASS-MDPs subsume deadlock-free VASS-MDPs and thus decidability (resp. undecidability) results carry over to the smaller (resp. larger) class. 2.4

Undecidability in the General Case

It was shown in [1] that the sure reachability problem is undecidable for (2dimensional) two player VASS. From this we can deduce that the sure reachability problem is undecidable for VASS-MDPs. We now present a similar proof to show the undecidability of the almost-sure reachability problem for VASS-MDPs. For all of our undecidability results we use reductions from the undecidable control-state reachability problem for Minsky machines. A Minsky machine is a tuple Q, T where Q is a finite set of states and T is a finite set of transitions manipulating two counters, say x1 and x2 . Each transition is a triple of the form

q, xi = 0?, q  (counter xi is tested for 0) or q, xi := xi + 1, q  (counter xi is incremented) or q, xi := xi − 1, q  (counter xi is decremented) where q, q  ∈ Q.

Qualitative Analysis of VASS-Induced MDPs

325

Fig. 2. Encoding q1 , x1 := x1 + 1, q2  and q3 , x2 := x2 − 1, q4  and q5 , x1 = 0?, q6 

Configurations of a Minsky machine are triples in Q × N × N. The transition relation ⇒ between configurations of the Minsky machine is then defined in the obvious way. Given an initial state qI and a final state qF , the controlstate reachability problem asks whether there exists a sequence of configurations

qI , 0, 0 ⇒ q1 , v1 , v1 ⇒ . . . ⇒ qk , vk , vk with qk = qF . This problem is known to be undecidable [16]. W.l.o.g. we assume that Minsky machines are deadlockfree and deterministic (i.e., each configuration has always a unique successor) and that the only transition leaving qF is of the form qF , x1 := x1 + 1, qF . We now show how to reduce the control-state reachability problem to the almost-sure and limit-sure reachability problems in deadlock-free VASS-MDPs. From a Minsky machine, we construct a deadlock-free 2-dimensional VASS-MDP for which the control states of Player 1 are exactly the control states of the Minsky machine. The encoding is presented in Fig. 2 where the circles (resp. squares) are the control states of Player 1 (resp. Player P), and for each edge the corresponding weight is 1. The state ⊥ is an absorbing state from which the unique outgoing transition is a self loop that does not affect the values of the counters. This encoding allows us to deduce our first result. Theorem 1. The sure, almost-sure and limit-sure (repeated) reachability problems are undecidable problems for 2-dimensional deadlock-free VASS-MDPs. In the special case of 1-dimensional VASS-MDPs, the sure and almost-sure reachability problems are decidable [7]. 2.5

Model-Checking µ-calculus on Single-Sided VASS

It is well-known that there is a strong connection between model-checking branching time logics and games, and in our case we have in fact undecidability results for simple reachability games played on a VASS and for the modelchecking of VASS with expressive branching-time logics [12]. However for this latter point, decidability can be regained by imposing some restrictions on the VASS structure [4] as we will now recall. We say that a VASS Q, T is (Q1 , Q2 )single-sided iff Q1 and Q2 represents a partition of the set of states Q such that for all transitions q, z, q  in T with q ∈ Q2 , we have z = 0; in other words only the transitions leaving a state from Q1 are allowed to change the values of the counters. In [4], it has been shown that, thanks to a reduction to games played on a single-sided VASS with parity objectives, a large fragment of the μ-calculus called Lsv μ has a decidable model-checking problem over single-sided

326

P.A. Abdulla et al.

VASS. The idea of this fragment is that the “always” operator  is guarded with a predicate enforcing the current control states to belong to Q2 . Formally, the syntax of Lsv μ for (Q1 , Q2 )-single-sided VASS is given by the following grammar: φ ::= q | X | φ ∧ φ | φ ∨ φ | ♦φ | Q2 ∧ φ | μX.φ | νX.φ, where Q2 stands for the formula q∈Q2 q and X belongs to a set of variables X . The semantics of Lsv μ is defined as usual: it associates to a formula φ and to an environment ε : X → 2C a subset of configurations φε . We use ε0 to denote the environment which assigns the empty set to any variable. Given an environment ε, a variable X ∈ X and a subset of configurations C, we use ε[X := C] to represent the environment ε which is equal to ε except on the variable X, where we have ε (X) = C. Finally the notation φ corresponds to the interpretation φε0 . The problem of model-checking single-sided VASS with Lsv μ can then be defined as follows: given a single-sided VASS Q, T , an initial configuration c0 and a formula φ of Lsv μ , do we have c0 ∈ φ? Theorem 2 [4]. Model-checking single-sided VASS wrt. Lsv μ is decidable.

3

Verification of P-VASS-MDPs

In [4] it is proved that parity games played on a single-sided deadlock-free VASS are decidable (this entails the decidability of model checking Lsv μ over singlesided VASS). We will see here that in the case of P-VASS-MDPs, in which only the probabilistic player can modify the counters, the decidability status depends on the presence of deadlocks in the system. 3.1

Undecidability in Presence of Deadlocks

We point out that the reduction presented in Fig. 2 to prove Theorem 1 does not carry over to P-VASS-MDPs, because in that construction both players have the ability to change the counter values. However, it is possible to perform a similar reduction leading to the undecidability of verification problems for PVASS-MDPs, the main difference being that we crucially exploit the fact that the P-VASS-MDP can contain deadlocks. We now explain the idea behind our encoding of Minsky machines into PVASS-MDPs. Intuitively, Player 1 chooses a transition of the Minsky machine to simulate, anticipating the modification of the counters values, and Player P is then in charge of performing the change. If Player 1 chooses a transition with a decrement and the accessed counter value is actually 0, then Player P will be in a deadlock state and consequently the desired control state will not be reached. Furthermore, if Player 1 decides to perform a zero-test when the counter value is strictly positive, then Player P is able to punish this choice by entering a deadlock state. Similarly to the proof of Theorem 1, Player P can test if the value of the counter is strictly greater than 0 by decrementing it. The encoding of the Minsky machine is presented in Fig. 3. Note that no outgoing edge of Player 1’s states changes the counter values. Furthermore, we see that Player P

Qualitative Analysis of VASS-Induced MDPs

327

Fig. 3. Encoding q1 , x1 := x1 + 1, q2  and q3 , x2 := x2 − 1, q4  and q5 , x1 = 0?, q6 

reaches the control state ⊥ if and only if Player 1 chooses to take a transition with a zero-test when the value of the tested counter is not equal to 0. Note that, with the encoding of the transition q3 , x2 := x2 − 1, q4 , when Player P is in the control state between q3 and q4 , it can be in a deadlock if the value of the second counter is not positive. In the sequel we will see that in P-VASS-MDP without deadlocks the sure reachability problem becomes decidable. From this encoding we deduce the following result. Theorem 3. The sure, almost sure and limit sure (repeated) reachability problems are undecidable for 2-dimensional P-VASS-MDPs. 3.2

Sure (repeated) Reachability in Deadlock-Free P-VASS-MDPs

Unlike in the case of general P-VASS-MDPs, we will see that the sure (repeated) reachability problem is decidable for deadlock-free P-VASS-MDPs. Let S =

Q, Q1 , QP , T, τ be a deadlock-free P-VASS-MDP, MS = (C, C1 , CP , →, p) its associated MDP and qF ∈ Q a control state. Note that because the P-VASSMDP S is deadlock free, Player P cannot take the play to a deadlock to avoid the control state qF , but he has to deal only with infinite plays. Since S is a P-VASS-MDP, the VASS Q, T is (QP , Q1 )-single-sided. In [1,18], it has been shown that control-state reachability games on deadlock-free single-sided VASS are decidable, and this result has been extended to parity games in [4]. This implies the decidability of sure (repeated) reachability in deadlock-free P-VASSMDPs. However, to obtain a generic way of verifying these systems, we construct a formula of Lsv μ that characterizes the sets of winning configurations and use then the result of Theorem 2. Let VSP be the set of configurations from which the answer to the sure reachability problem (with qF as state to be reached) is negative, i.e., VSP = {c ∈ C | σ ∈ Σ s.t. Plays(MS , c, σ) ⊆ ♦qF } and similarly let WSP = {c ∈ C | σ ∈ Σ s.t. Plays(MS , c, σ) ⊆ ♦qF }. The next sv lemma relates  QP corresponds to  these two sets with a formula of Lμ (where the formula q∈QP and Q1 corresponds to the formula q∈Q1 q).  Lemma 1. – VSP = νX.( q∈Q\{qF } q) ∧ (Q1 ∨ ♦X) ∧ (QP ∨ (Q1 ∧ X)).   – WSP = μY.νX. ( q∈Q\{qF } q) ∧ (Q1 ∨ ♦X) ∧ (QP ∨ (Q1 ∧ X)) ∨ (qF ∧ QP ∧  ♦Y ) ∨ (qF ∧ Q1 ∧ Y ) 

328

P.A. Abdulla et al.

We use (QP ∨ (Q1 ∧ X)) instead of (QP ∨ X) so that the formulae are in the guarded fragment of the μ-calculus. Since the two formulae belong to Lsv μ for the (QP , Q1 )-single-sided VASS S, decidability follows directly from Theorem 2. Theorem 4. The sure reachability and repeated reachability problem are decidable for deadlock free P-VASS-MDPs. 3.3

Almost-Sure and Limit-Sure Reachability in Deadlock-Free P-VASS-MDPs

We have seen that, unlike for the general case, the sure reachability and sure repeated reachability problems are decidable for deadlock free P-VASS-MDPs, with deadlock freeness being necessary to obtain decidability. For the corresponding almost-sure and limit-sure problems we now show undecidability, again using a reduction from the reachability problem for two counter Minsky machines, as shown in Fig. 4. The main difference with the construction used for the proof of Theorem 3 lies in the addition of a self-loop in the encoding of the transitions for decrementing a counter, in order to avoid deadlocks. If Player 1, from a configuration q3 , v , chooses the transition q3 , x2 := x2 − 1, q4 which decrements the second counter, then the probabilistic state with the self-loop is entered, and there are two possible cases: if v(2) > 0 then the probability of staying forever in this loop is 0 and the probability of eventually going to state q4 is 1; on the other hand, if v(2) = 0 then the probability of staying forever in the self-loop is 1, since the other transition that leaves the state of Player P and which performs the decrement on the second counter effectively is not available. Note that such a construction does not hold in the case of sure reachability, because the path that stays forever in the loop is a valid path.

Fig. 4. Encoding q1 , x1 := x1 + 1, q2  and q3 , x2 := x2 − 1, q4  and q5 , x1 = 0?, q6 

This allows us to deduce the following result for deadlock free P-VASS-MDPs. Theorem 5. The almost-sure and limit-sure (repeated) reachability problems are undecidable for 2-dimensional deadlock-free P-VASS-MDPs.

Qualitative Analysis of VASS-Induced MDPs

4

329

Verification of 1-VASS-MDPs

In this section, we will provide decidability results for the subclass of 1-VASSMDPs. As for deadlock-free P-VASS-MDPs, the proofs for sure and almostsure problems use the decidability of Lsv μ over single-sided VASS, whereas the technique used to show decidability of limit-sure reachability is different. 4.1

Sure Problems in 1-VASS-MDPs

First we show that, unlike for P-VASS-MDPs, deadlocks do not matter for 1VASS-MDPs. The idea is that in this case, if the deadlock is in a probabilistic configuration, it means that there is no outgoing edge (because of the property of 1-VASS-MDPs), and hence one can add an edge to a new absorbing state, and the same can be done for the states of Player 1. Such a construction does not work for P-VASS-MDPs, because in that case deadlocks in probabilistic configurations may depend on the counter values, and not just on the current control-state. Lemma 2. The sure (resp. almost sure, resp. limit sure) (repeated) reachability problem for 1-VASS-MDPs reduces to the sure (resp. almost sure, resp. limitsure) (repeated) reachability problem for deadlock-free 1-VASS-MDPs. Hence in the sequel we will consider only deadlock-free 1-VASS-MDPs. Let S = Q, Q1 , QP , T, τ be a deadlock-free 1-VASS-MDP. For what concerns the sure (repeated) reachability problems we can directly reuse the results from Lemma 1 and then show that the complement formulae of the ones expressed in this lemma belong to Lsv μ for the (Q1 , QP )-single-sided VASS Q, T (in fact the correctness of these two lemmas did not depend on the fact that we were considering P-VASS-MDPs). Theorem 2 allows us to retrieve the decidability results already expressed in [18] (for sure reachability) and [4] (for sure repeated reachability). Theorem 6. The sure (repeated) reachability problem is decidable for 1-VASSMDPs. 4.2

Almost-Sure Problems in 1-VASS-MDPs

We now move to the case of almost-sure problems in 1-VASS-MDPs. We consider a deadlock free 1-VASS-MDP S = Q, Q1 , QP , T, τ and its associated MDP MS = C, C1 , CP , →, p . We will see that, unlike for P-VASS-MDPs, it is here 1 also possible to characterize by formulae of Lsv μ the two following sets: VAS = 1 {c ∈ C | ∃σ ∈ Σ such that P(MS , c, σ, ♦qF ) = 1} and WAS = {c ∈ C | ∃σ ∈ Σ such that P(MS , c, σ, ♦qF ) = 1}, i.e. the set of configurations from which Player 1 has a strategy to reach the control state qF , respectively to visit infinitely often qF , with probability 1. We begin with introducing the following formula of Lsv μ based on the variables X and Y : InvPre(X, Y ) = (Q1 ∧ ♦(X ∧ Y )) ∨ (♦Y ∧ QP ∧ X). Note that

330

P.A. Abdulla et al.

InvPre(X, Y ) is a formula of Lsv μ for the (Q1 , QP )-single-sided VASS Q, T . Intuitively, this formula represents the set of configurations from which (i) Player 1 can make a transition to the set represented by the intersection of the sets characterized by the variables X and Y and (ii) Player P can make a transition to the set Y and cannot avoid making a transition to the set X. 1 can be characterized Almost Sure Reachability. We will now prove that VAS sv by the following formula of Lμ : νX.μY.(qF ∨ InvPre(X, Y )). Note that a similar result exists for finite-state MDPs, see e.g. [9]; this result in general does not extend to infinite-state MDPs, but in the case of VASS-MDPs it can be applied. Before proving this we need some intermediate results.    We denoteby E the set νX.μY. qF ∨ InvPre(X, Y ) ε0 . Since νX.μY. qF ∨ InvPre(X, Y ) is a formula of Lsv μ interpreted over the single-sided VASS Q, T , we can show that E is an upward-closed set. We now need another lemma which states that there exists N ∈ N and a strategy for Player 1 such that, from any configuration of E, Player 1 can reach the control state qF in less than N steps and Player P cannot take the play outside of E. The   fact that we can bound the number of steps is crucial to show that νX.μY. qF ∨ InvPre(X, Y ) ε0 is 1 . For infinite-state MDPs where this property does not hold, our equal to VAS techniques do not apply.

Lemma 3. There exists N ∈ N and a strategy σ of Player 1 such that for all c ∈ E, there exists a play c · c1 · c2 · . . . in Plays(MS , c, σ) satisfying the three following properties: (1) there exists 0 ≤ i ≤ N such that ci ∈ qF ; (2) for all 0 ≤ j ≤ i, cj ∈ E; (3) for all 0 ≤ j ≤ i, if cj ∈ CP then for all c ∈ C such that cj → c , we have c ∈ E. 1 This previous lemma allows us to characterize VAS with a formula of Lsv μ . The proof of the following result uses the fact that the number of steps is bounded, and also the fact that the sets described by closed Lsv μ formulae are upwardclosed. This makes the fixpoint iteration terminate in a finite number of steps. 1 Lemma 4. VAS = νX.μY.(qF ∨ InvPre(X, Y )). 1 Since Q, T is (Q1 , QP )-single-sided and since the formula associated to VAS sv belongs to Lμ , from Theorem 2 we deduce the following theorem.

Theorem 7. The almost-sure reachability problem is decidable for 1-VASSMDPs. Almost Sure Repeated Reachability. For the case of almost sure repeated reachability we reuse the previously introduced formula InvPre(X, Y ). We can perform a reasoning similar to the previous ones and provide a characterization 1 . of the set WAS 1 Lemma 5. WAS = νX.InvPre(X, μY.(qF ∨ InvPre(X, Y ))).

Qualitative Analysis of VASS-Induced MDPs

331

As previously, this allows us to deduce the decidability of the almost sure repeated reachability problem for 1-VASS-MDP. Theorem 8. The almost sure repeated reachability problem is decidable for 1VASS-MDPs. 4.3

Limit-Sure Reachability in 1-VASS-MDP

We consider a slightly more general version of the limit-sure reachability problem with a set X ⊆ Q of target states instead of a single state qF , i.e., the standard case corresponds to X = {qF }.  We extend the set of natural numbers N to N∗ = N {∗} by adding an element ∗ ∈ / N with ∗+j = ∗−j = ∗ and j < ∗ for all j ∈ N. We consider then the set of vectors Nd∗ . The projection of a vector v in Nd by eliminating components that are indexed by a natural number k is defined by projk (v)(i) = v(i) if i = k and projk (v)(i) = ∗ otherwise Let Qc represent control-states which are indexed by a color. The coloring functions coli : Q → Qc create colored copies of control-states by coli (q) = qi . Given a 1-VASS-MDP S = Q, Q1 , QP , T, τ of dimension d, an index k ≤ d and a color i, the colored projection is defined as: P rojk (S, d, i) = coli (Q), coli (Q1 ), coli (QP ), projk,i (T ), τk,i where projk,i (T ) = {projk,i (t)|t ∈ T } is the projection of the set of transitions T and projk,i (t) = coli (x), projk (z), coli (y) is the projection of transition t =

x, z, y obtained by removing component k and coloringthe states x and y with color i. The transition weights carry over, i.e., τk,i (t ) = {τ (t) | projk,i (t) = t }. We define the functions state : Q × Nd∗ → Q and count : Q × Nd∗ → Nd∗ s.t for a configuration ci = q, v , where q ∈ Q and v ∈ Nd∗ we have that state(q, v) = q and count(q, v) = v. For any two configurations c1 and c2 , we write c1 ≺ c2 to denote that state(c1 ) = state(c2 ), and there exists a nonempty set of indexes I where for every i ∈ I , count(c1 )(i) < count(c2 )(i), whereas for every index j∈ / I, 0 < j ≤ d, count(c1 )(j) = count(c2 )(j). Algorithm 1 reduces the dimension of the limit-sure reachability problem for 1-VASS-MDP by a construction resembling the Karp-Miller tree [15]. It takes as input a 1-VASS-MDP S of some dimension d > 0 with a set of target states X. It outputs a new 1-VASS-MDP S  of dimension d − 1 and a new set of target states X  such that MS can limit-surely reach X iff MS  can limit-surely reach X  . In particular, in the base case where d−1 = 0, the new system S  has dimension zero and thus induces a finite-state MDP MS  , for which limit-sure reachability of X  coincides with almost-sure reachability of X  , which is known to be decidable in polynomial time. Algorithm 1 starts by exploring all branches of the computation tree of S (and adding them to S  as the so-called initial uncolored part) until it encounters a configuration that is either (1) equal to, or (2) strictly larger than a configuration encountered previously on the same branch. In case (1) it just adds a back loop to the point where the configuration was encountered

332

P.A. Abdulla et al.

previously. In case (2), it adds a modified copy of S (identified by a unique color) to S  . This so-called colored subsystem is similar to S except that those counters that have strictly increased along the branch are removed. The intuition is that these counters could be pumped to arbitrarily high values and thus present no obstacle to reaching the target. Since the initial uncolored part is necessarily finite (by Dickson’s Lemma) and each of the finitely many colored subsystems only has dimension d − 1 (since a counter is removed; possibly a different one in different colored subsystems), the resulting 1-VASS-MDP S  has dimension d − 1. The set of target states X  is defined as the union of all appearances of states in X in the uncolored part, plus all colored copies of states from X in the colored subsystems.

Algorithm 1. Reducing the dimension of the limit-sure reachability problem. Require: S = Q, Q1 , QP , T, τ  1-VASS-MDP, dimension d > 0, c0 = q0 , v ∈ Q × Nd X ⊆ Q - set of target states  Ensure: S  = Q , Q1 , QP , T  , τ  ; c0 = q0 , 0; X  ⊆ Q ; λ : Q → ((Q Qc ) × Nd∗ ) 1: Q ← ∅; Q1 ← ∅; QP ← ∅; T  ← ∅; τ  ← ∅; 2: new(q  ); q0 ← q  ; λ(q  ) ← c0 ; Q ← {q  }; i ← 0 3: if state(λ(q  )) ∈ Q1 then Q1 ← {q  } else QP ← {q  } 4: ToExplore ← {q  } 5: while ToExplore = ∅ do 6: Pick and remove a q ∈ ToExplore 7: if ∃q  . q  is previously on the same branch as q and λ(q  ) ≺ λ(q) then 8: get indexes I in which the counter is increasing 9: pick and remove the first index k from I 10: i ← i + 1; // increase color index 11: new(q  ); 12: λ(q  ) ← coli (state(λ(q))), projk (count(λ(q)))   13: if state(λ(q)) ∈ Q1 then Q1 ← Q1 {q  } else QP ← QP {q  }      ) = 1; 14: T ← T {q, 0, q }; τ (q, 0, q   15: Q1 ← Q1  coli (Q1 ); QP ← QP coli (QP ); T  ← T  projk,i (T ); 16: X  ← X  coli (X); τ  ← τ  ∪ τk,i 17: else 18: for every t = x, z, y in T such that t ∈ En(λ(q)) do 19: if ∃q  . q  ispreviously on the same branch as q and t(λ(q)) = λ(q  ) then 20: T  ← T  {q, z, q  }; 21: else λ(q  ) ← t(λ(q)) 22: new(q  );    23: T ← T {q, z, q  }; τ  (q, z, q  ) ← τ (t)  24: if state(λ(q  )) ∈ Q1 then Q1 ← Q1 {q  } else QP ← QP {q  }     25: if state(λ(q )) ∈ X then X  ← X {q } 26: T oExplore ← T oExplore {q  } 27: end if 28: end for 29: end if 30: end while

Qualitative Analysis of VASS-Induced MDPs

333

Table 1. Decidability of verification problems for P-VASS-MDP, deadlock-free PVASS-MDP and 1-VASS-MDP. A  stands for decidable and a × for undecidable. P-VASS-MDP df P-VASS-MDP 1-VASS-MDP Sure reachability

× (Theorem 3)  (Theorem 4)

 (Theorem 6)

Almost-sure reachability

× (Theorem 3) × (Theorem 5)

 (Theorem 7)

Limit-sure reachability

× (Theorem 3) × (Theorem 5)

 (Theorem 9)

Sure repeated reachability

× (Theorem 3)  (Theorem 4)

 (Theorem 6)

Almost-sure repeated reachability × (Theorem 3) × (Theorem 5)

 (Theorem 8)

Limit-sure repeated reachability

× (Theorem 3) × (Theorem 5)

Open

By Dickson’s Lemma, the conditions on line 7 or line 19 of the algorithm must eventually hold on every branch of the explored computation tree. Thus, it will terminate. Lemma 6. Algorithm 1 terminates. The next lemma states the correctness of Algorithm 1. Let S =

Q, Q1 , QP , T, τ be 1-VASS-MDP of dimension d > 0 with initial configuration c0 = q0 , v and X ⊆ Q a set of target states. Let S  = Q , Q1 , QP , T  , τ  with initial configuration c0 = q0 , 0 and set of target states X  ⊆ Q be the (d − 1) dimensional 1-VASS-MDP produced by Algorithm 1. As described above we have the following relation between these two systems. Lemma 7. P+ (MS , c0 , ♦X) = 1 iff P+ (MS  , c0 , ♦X  ) = 1. By applying the result of the previous lemma iteratively until we obtain a finite-state MDP, we can deduce the following theorem. Theorem 9. The limit-sure reachability problem for 1-VASS-MDP is decidable.

5

Conclusion and Future Work

Table 1 summarizes our results on the decidability of verification problems for subclasses of VASS-MDP. The exact complexity of most problems is still open. Algorithm 1 relies on Dickson’s Lemma for termination, and the algorithm deciding the model-checking problem of Theorem 2 additionally uses the Valk-Jantzen construction repeatedly. However, all these problems are at least as hard as control-state reachability in VASS, and thus EXPSPACE-hard [12]. The decidability of the limit-sure repeated reachability problem for 1-VASSMDP is open. A hint of its difficulty is given by the fact that there are instances where the property holds even though a small chance of reaching a deadlock cannot be avoided from any reachable configuration. In particular, a solution would require an analysis of the long-run behavior of multi-dimensional random walks induced by probabilistic VASS. However, these may exhibit strange nonregular behaviors for dimensions ≥ 3, as described in [8] (Sect. 5).

334

P.A. Abdulla et al.

References 1. Abdulla, P.A., Bouajjani, A., d’Orso, J.: Monotonic and downward closed games. J. Logic Comput. 18(1), 153–169 (2008) 2. Abdulla, P.A., Ciobanu, R., Mayr, R., Sangnier, A., Sproston, J.: Qualitative analysis of VASS-induced MDPs. CoRR (2015). abs/1512.08824 3. Abdulla, P.A., Henda, N.B., Mayr, R.: Decisive Markov chains. Logical Meth. Comput. Sci. 3(4) (2007) 4. Abdulla, P.A., Mayr, R., Sangnier, A., Sproston, J.: Solving parity games on integer vectors. In: D’Argenio, P.R., Melgratti, H. (eds.) CONCUR 2013 – Concurrency Theory. LNCS, vol. 8052, pp. 106–120. Springer, Heidelberg (2013) 5. Baier, C., Katoen, J.-P.: Principles of Model Checking. MIT Press, Cambridge (2008) 6. Br´ azdil, T., Brozek, V., Etessami, K.: One-counter stochastic games. In: FSTTCS 2010, LIPIcs, vol. 8, pp. 108–119. Schloss Dagstuhl - Leibniz-Zentrum f¨ ur Informatik (2010) 7. Br´ azdil, T., Brozek, V., Etessami, K., Kuˇcera, A., Wojtczak, D.: One-counter Markov decision processes. In: SODA 2010, pp. 863–874. SIAM (2010) 8. Br´ azdil, T., Kiefer, S., Kuˇcera, A., Novotn´ y, P.: Long-run average behaviour of probabilistic vector addition systems. In: LICS 2015, pp. 44–55. IEEE (2015) 9. Chatterjee, K., de Alfaro, L., Faella, M., Legay, A.: Qualitative logics and equivalences for probabilistic systems. Logical Meth. Comput. Sci. 5(2) (2009) 10. Chatterjee, K., Jurdzi´ nski, M., Henzinger, T.A.: Simple stochastic parity games. In: Baaz, M., Makowsky, J.A. (eds.) CSL 2003. LNCS, vol. 2803, pp. 100–113. Springer, Heidelberg (2003) 11. Condon, A.: The complexity of stochastic games. Inf. Comput. 96(2), 203–224 (1992) 12. Esparza, J., Nielsen, M.: Decidability issues for Petri nets - a survey. Bull. EATCS 52, 244–262 (1994) 13. Etessami, K., Yannakakis, M.: Recursive Markov decision processes and recursive stochastic games. In: Caires, L., Italiano, G.F., Monteiro, L., Palamidessi, C., Yung, M. (eds.) ICALP 2005. LNCS, vol. 3580, pp. 891–903. Springer, Heidelberg (2005) 14. Filar, J., Vrieze, K.: Competitive Markov Decision Processes. Springer, New York (1997) 15. Karp, R., Miller, R.: Parallel program schemata. J. Comput. Syst. Sci. 3(2), 147– 195 (1969) 16. Minsky, M.L.: Computation: Finite and Infinite Machines. Prentice-Hall, Upper Saddle River (1967) 17. Puterman, M.L.: Markov Decision Processes. Wiley, New York (1994) 18. Raskin, J.-F., Samuelides, M., Begin, L.V.: Games for counting abstractions. In: AVoCS 2004, Electronic Notes in Theoretical Computer Science, vol. 128(6), pp. 69–85 (2005) 19. Shapley, L.S.: Stochastic games. Proc. Nat. Acad. Sci. 39(10), 1095–1100 (1953)

Metric Temporal Logic with Counting Shankara Narayanan Krishna1(B) , Khushraj Madnani1 , and Paritosh K. Pandya2 1

Department of Computer Science and Engineering, IIT Bombay, Mumbai 400 076, India {krishnas,khushraj}@cse.iitb.ac.in 2 School of Technology and Computer Science, Tata Institute of Fundamental Research, Mumbai 400 005, India [email protected]

Ability to count number of occurrences of events within a specified time interval is very useful in specification of resource bounded real time computation. In this paper, we study an extension of Metric Temporal Logic (MTL) with two different counting modalities called C and UT (until with threshold), which enhance the expressive power of MTL in orthogonal fashion. We confine ourselves only to the future fragment of MTL interpreted in a pointwise manner over finite timed words. We provide a comprehensive study of the expressive power of logic CTMTL and its fragments using the technique of EF games extended with suitable counting moves. Finally, as our main result, we establish the decidability of CTMTL by giving an equisatisfiable reduction from CTMTL to MTL. The reduction provides one more example of the use of temporal projections with oversampling introduced earlier for proving decidability. Our reduction also implies that MITL extended with C and UT modalities is elementarily decidable.

1

Introduction

Temporal logics provide constructs to specify qualitative ordering between events in time. But real time logics have the ability to specify quantitative timing constraints between events. Metric Temporal Logic MTL is amongst the best studied of real time logics. Its principal modality a UI b states that an event b should occur in future within a time distance lying within interval I. Moreover, a should hold continuously till then. In many situations, especially those dealing with resource bounded computation, the ability to count the number of occurrences of events becomes important. In this paper, we consider an extension of MTL with two counting modalities C and UT (until threshold) which provide differing abilities to specify constraints on counts on events in time intervals. The resulting logic is called CTMTL. φ states that the number of times formula φ holds in time interval Modality C≥n I I (measured relative to current time point) is at least n. This is a mild generalization of C≥n (0,1) φ modality studied by Rabinovich [2] in context of continuous time MTL. The UT modality φ UI,#κ≥n ψ is like MTL until but it additionally states that the number of time formula κ holds between now and time point c Springer-Verlag Berlin Heidelberg 2016  B. Jacobs and C. L¨ oding (Eds.): FOSSACS 2016, LNCS 9634, pp. 335–352, 2016. DOI: 10.1007/978-3-662-49630-5 20

336

S.N. Krishna et al.

where ψ holds is at least n. Thus it extends U to simultaneously specify constraint on time and count of subformula. Constraining U by count of subformula was already explored for untimed LTL by Laroussini et al. [7]. But the combination of timing and counting seems new. The following example illustrates the use of these modalities. Example 1. We specify some constraints to be monitored by exercise bicycle electronics. – Two minutes after the start of exercise, the heartbeat (number of pulses in next 60 seconds) should be between 90 and 120. This can be stated as i, ρ, j |= ϕ2 , tj − ti ∈ I, and ρ, k |= ϕ1 ∀ i < k < j ρ satisfies ϕ denoted ρ |= ϕ iff ρ, 1 |= ϕ. Let L(ϕ) = {ρ | ρ, 1 |= ϕ} denote the language of a MTL formula ϕ. Two formulae ϕ and φ are said to be equivalent denoted as ϕ ≡ φ iff L(ϕ) = L(φ). Additional temporal connectives are defined in the standard way: we have the constrained future eventuality operator ♦I a ≡ trueUI a and its dual I a ≡ ¬♦I ¬a. We also define the next operator as OI φ ≡ w ⊥UI φ. Weak versions of operators are defined as ♦w I a = a ∨ ♦I a, I a ≡ a ∧ I a, b ≡ b ∨ [a ∧ (aU b)] if 0 ∈ I, and [a ∧ (aU b)] if 0 ∈ / I. aUw I I I

338

S.N. Krishna et al.

Theorem 1 [9]. Satisfiability checking of MTL is decidable over finite timed words with non-primitive-recursive complexity. Metric Temporal Logic with Counting (CTMTL). We denote by CTMTL the logic obtained by extending MTL with the ability to count, by endowing it with two counting modalities C as well as UT. Syntax of CTMTL: ϕ ::= a(∈ Σ)|true|ϕ ∧ ϕ|¬ϕ|ϕ|C≥n I ϕ|ϕUI,η ϕ, where I ∈ Iν, n ∈ N ∪ {0} and η is a threshold formula of the form #ϕ ≥ n or #ϕ < n. The counting modality C≥n I ϕ is called the C modality, while ϕUI,η ϕ is called the UT modality. Let ρ = (σ, τ ) ∈ T Σ ∗ , i, j ∈ dom(ρ). Define N ρ [i, I](ϕ) = {k ∈ dom(ρ) | tk ∈ ti + I ∧ ρ, k |= ϕ}, and ρ[i, j](ϕ) = {k ∈ dom(ρ) | i < k < j ∧ ρ, k |= ϕ}. Denote by |N ρ [i, I](ϕ)| and |ρ[i, j](ϕ)| respectively, the cardinality of N ρ [i, I](ϕ) and ρ[i, j](ϕ). |N ρ [i, I](ϕ)| is the number of points in ρ that lie in the interval ti +I, and which satisfy ϕ, while |ρ[i, j](ϕ)| is the number of points lying between ρ i and j which satisfy ϕ. Define ρ, i |= C≥n I ϕ iff |N [i, I](ϕ)| ≥ n. Likewise, ρ, i |= ϕ1 UI,#ϕ≥n ϕ2 iff ∃j>i, ρ, j |= ϕ2 , tj − ti ∈ I, and ρ, k |= ϕ1 , ∀i < k < j and |ρ[i, j](ϕ)| ≥ n. Remark : The classical until operator of MTL is captured in CTMTL since and #ϕ ∼ n for ∼∈ {≤, , =} ϕUI ψ ≡ ϕUI,#true≥0 ψ. We can express C∼n I ≥n ≥n+1 >n ϕ ≡ ¬C ϕ, C ϕ ≡ C≥n+1 ϕ, C≤n ϕ and in CTMTL since C n ≡ #ϕ ≥ n + 1, #ϕ ≤ n ≡ ¬(#ϕ > n + 1). It can be shown that boolean combinations of threshold formulae are also expressible in CTMTL (see [6]). Thus, aU(1,2),(#d=3∧#C ip ), and pebbles a set of points that are also present between jp and jp . Situation 2: Starting from (i1 , j1 ) with time stamps (0,0), if the Spoiler chooses a U(0,1),#a∼c move and lands up at some point between x1 and y1 , Duplicator will play copy-cat and achieve an identical configuration. Consider the case when Spoiler lands up at y1 2 . In response, Duplicator moves to y1 . From configuration (i2 , j2 ) with time stamps (y1 , y1 ), consider the case when Spoiler initiates a U(1,2),#a∼c and moves to z2 = 1.8 + < 2. In response, Duplicator moves to the point z2 = 2.1 > 2. A pebble is kept at the inbetween positions x2 , x2 respectively in ρ1 , ρ2 . When Spoiler picks the pebble in Duplicator’s word, then we obtain the configuration (i3 , j3 ) with time stamps (x2 , x2 ). If Spoiler does not get into the counting part/until part, the configuration obtained has time stamps (z2 , z2 ), with the lag of one segment (seg(i3 ) = 1, seg(j3 ) = 2, seg(j3 )seg(i3 )=1). We show in [6] that from (i3 , j3 ) with time stamps either (x2 , x2 ) or (z2 , z2 ), Duplicator can either achieve an identical configuration, or achieve a configuration with a lag of one segment. From situations (1), (2) in Proposition 1, we know that either Duplicator achieves an identical configuration, in which case there is no segment lag, or there is a lag of at most one segment. The length of the words are lnk + nl = K. If Spoiler always chooses bounded intervals (of length ≤ l), then Duplicator respects his segment lag of 1, and the maximum number of segments that can be explored in either word is at most nl < K. In this case, after p rounds, there are at least K − pl ≥ nlk + nl − pl ≥ (n − p)(l + 1) segments to the right of ρ1 and K −pl +1 segments to the right of ρ2 . If Spoiler chooses an unbounded interval in any round, then Duplicator can either enforce an identical configuration in both situations 1 and 2, or obtain one of the configurations with time stamps (p, f (p)), f (p) = x2 , or (z2 , x2 ) or (x2 , x2 ), from where it is known that Duplicator wins. Lemma 2. MTL ⊂ C(0,1) MTL ⊂ C(0,u) MTL Proof. We show that the formula ϕ = C=2 (0,1) a ∈ C(0,1) MTL cannot be expressed

in MTL. Likewise, the formula ϕ = C≥2 (0,2) a ∈ C(0,u) MTL cannot be expressed in C(0,1) MTL. A detailed proof of these are given in [6]. Lemma 3. (i) C(0,u) MTL ⊂ TMTL = C(0,u) TMTL = C(0,1) TMTL and (ii) C(0,u) MTL ⊂ CMTL. 2

The argument when Spoiler lands up at x1 or a point in between x1 , y1 is exactly the same.

Metric Temporal Logic with Counting

345

Proof. (i) The first containment as well as the last two equalities follows from the fact that the counting modality C≥n 0,j ϕ of C(0,u) MTL can be written in TMTL as ♦ 0,j ,#ϕ≥n true. The strict containment of C(0,u) MTL then follows from Lemma 4. (ii) We know that C(0,u) MTL ⊆ CMTL. This along with (i) and Lemma 1 gives the strict containment. Lemma 4. TMTL − CMTL = ∅ Proof. Consider the formula ϕ = ♦(0,1),#a≥3 b ∈ TMTL. We show that for any choice of n rounds and k pebbles, we can find two words ρ1 , ρ2 such that ρ2 |= ρ2 . The words can be seen in Fig. 2 and the details in [6]. ϕ, ρ1  ϕ, but ρ1 ≡CMTL n,k

Fig. 2. The red square represents a, the block of blue lines represents a block of b’s. There are 3 a’s in each unit interval of both ρ1 and ρ2 . The difference is that ρ1 has 3 blocks of b’s in each unit interval, while ρ2 has 4 blocks of b’s in each unit interval except the last. Clearly, ρ2 |= ϕ, ρ1  ϕ (Color figure online).

4

Satisfiability Checking of Counting Logics

In this section, we show that CTMTL has a decidable satisfiability checking. For this, given a formula in CTMTL we synthesize an equisatisfiable formula in MTL, and use the decidability of MTL. We start discussing some preliminaries. Let Σ, X be finite sets of propositions such that Σ ∩ X = ∅. 1. (Σ, X)-simple extensions. A (Σ, X)-simple extension is a timed word ρ = (σ  , τ  ) over X ∪ Σ such that at any point i ∈ dom(ρ ), σi ∩ Σ = ∅. For Σ = {a, b}, X = {c, d}, ({a}, 0.2)({b, c, d}, 0.3)({b, d}, 1.1) is a (Σ, X)-simple extension. However, ({a}, 0.2)({c, d}, 0.3)({b, d}, 1.1) is not. 2. Simple Projections. Consider a (Σ, X)-simple extension ρ. We define the simple projection of ρ with respect to X, denoted ρ \ X as the word obtained by erasing the symbols of X from each σi . Note that dom(ρ) = dom(ρ \ X). For example, if Σ = {a, c}, X = {b}, and ρ = ({a, b, c}, 0.2)({b, c}, 1)({c}, 1.3), then ρ \ X = ({a, c}, 0.2)({c}, 1)({c}, 1.3). ρ \ X is thus, a timed word over Σ. If the underlying word ρ is not a (Σ, X)-simple extension, then ρ \ X is undefined. 3. (Σ, X)-oversampled behaviours. A (Σ, X)-oversampled behaviour is a timed  word ρ = (σ  , τ  ) over X ∪ Σ, such that σ1 ∩ Σ = ∅ and σ|dom(ρ  )| ∩ Σ = ∅. Oversampled behaviours are more general than simple extensions since they allow occurrences of new points in between the first and the last position. These new points are called oversampled points.

346

S.N. Krishna et al.

All other points are called action points. For Σ = {a, b}, X = {c, d}, ({a}, 0.2)({c, d}, 0.3)({a, b}, 0.7)({b, d}, 1.1) is a (Σ, X)-oversampled behaviour, while ({a}, 0.2)({c, d}, 0.3)({c}, 1.1) is not. 4. Oversampled Projections. Given a (Σ, X)-oversampled behaviour ρ = (σ  , τ  ), the oversampled projection of ρ with respect to Σ, denoted ρ ↓ X is defined as the timed word obtained by deleting the oversampled points, and then erasing the symbols of X from the action points. ρ=ρ ↓ X is a timed word over Σ. A temporal projection is either a simple projection or an oversampled projection. We now define equisatisfiability modulo temporal projections. Given MTL formulae ψ and φ, we say that φ is equisatisfiable to ψ modulo temporal projections iff there exist disjoint sets X, Σ such that (1) φ is over Σ, and ψ over Σ ∪ X, (2) For any timed word ρ over Σ such that ρ |= φ, there exists a timed word ρ such that ρ |= ψ, and ρ is a temporal projection of ρ with respect to X, (3) For any behaviour ρ over Σ ∪ X, if ρ |= ψ then the temporal projection ρ of ρ with respect to X is well defined and ρ |= φ. If the temporal projection used above is a simple projection, we call it equisatisfiability modulo simple projections and denote it by φ = ∃X.ψ. If the projection in the above definition is an oversampled projection, then it is called equisatisfiability modulo oversampled projections and is denoted φ ≡ ∃ ↓ X.ψ. Equisatisfiability modulo simple projections are studied extensively [5, 11, 13]. It can be seen that if ϕ1 = ∃X1 .ψ1 and ϕ2 = ∃X2 .ψ2 , with X1 , X2 disjoint, then ϕ1 ∧ ϕ2 = ∃(X1 ∪ X2 ).(ψ1 ∧ ψ2 ) [8]. As in the case of simple projections, equisatisfiability modulo oversampled projections are also closed under conjunctions when one considers the relativized formulae. For example, consider a formula ϕ = (0,1) a over Σ = {a, d}. Let ψ1 = (0,1) (a ∨ b) ∧ ♦(0,1) (b ∧ ¬a) be a formula over the extended alphabet {a, b, d} and ψ2 = (c ↔ (0,1) a) ∧ c over the extended alphabet {a, c, d}. Note that ϕ = ∃ ↓ {b}.ψ1 and ϕ = ∃ ↓ {c}.ψ2 but ϕ∧ϕ = ∃ ↓ {b, c}.(ψ1 ∧ψ2 ) as the left hand side evaluates to ϕ which is satisfiable while the right hand side is unsatisfiable. This is due to the presence of a non-action point where only b holds. But this can easily be fixed by relativizing all the formulae over their respective action points. ψ1 is relativized as λ1 = (0,1) (act1 → (a∨b))∧♦(0,1) (act1 ∧b∧¬a) and ψ2 is relativized as λ2 = (act2 → (c ↔ (0,1) (act2 → a))) ∧ act2 ∧ c where act1 = b ∨ d ∨ a and act2 = a ∨ c ∨ d. Now, ϕ ∧ ϕ = ∃ ↓ {b, c}.(λ1 ∧ λ2 ). The relativized forms of ψ1 , ψ2 are called their Oversampled Normal Forms with respect to Σ and denoted ON FΣ (ψ1 ) and ON FΣ (ψ2 ). Then it can be seen that ϕ1 ∧ ϕ2 = ∃ ↓ {b, d}.[ON FΣ (ψ1 ) ∧ ON FΣ (ψ2 )], and ϕ1 = ∃ ↓ {b}.ON FΣ (ψ1 ), ϕ2 = ∃ ↓ {d}.ON FΣ (ψ2 ). The formal definition of ON FΣ (ϕ) for a formula ϕ over Σ ∪ X can be found in [6]. Equisatisfiability modulo oversampled projections were first studied in [4] to eliminate non-punctual past from MTL over timed words. We use equisatifiability modulo simple projections to eliminate the C modality and oversampled projections to eliminate the UT modality from CTMTL.

Metric Temporal Logic with Counting

347

Elimination of Counting Modalities from CTMTL. In this section, we show how to eliminate the counting constraints from CTMTL over strictly monotonic timed words. This can be extended to weakly monotonic timed words. Given any CTMTL formula ϕ over Σ, we “flatten” the C, UT modalities of ϕ and obtain a flattened formula. As an example, consider the formula ϕ = ≥2 aU[0,3] (c ∧ C=1 e)≥1 C(0,1) e]). Replacing the counting modalities (2,3) dU(0,1),#(d∧C=1 (0,1) with fresh witness propositions w1 , w2 , we obtain ϕf lat = [aU[0,3] (c ∧ w1 )] ∧ T w where T = T1 ∧ T2 ∧ T3 ∧ T4 , with T1 = w [w1 ↔ C=1 (2,3) w2 ], T2 =  [w2 ↔

≥2 w dU(0,1),#w4 ≥1 w3 ]], T3 = w [w4 ↔ (d∧C=1 (0,1) e)], and T4 =  [w3 ↔ C(0,1) e]. Each temporal projection Ti obtained after flattening contains either a C modality or a UT modality. In the following, we now show how to obtain equisatisfiable MTL formulae corresponding to each temporal projection.

Lemma 5. The formula C≥n l,∞) b is equivalent to MTL formula F l,∞) (b ∧ F (b ∧ . . . F b))). We now outline the steps followed to obtain an equisatisfiable formula in MTL, assuming C≥n l,∞) b modalities have been eliminated using Lemma 5. 1. Flattening: Flatten χ obtaining χf lat over Σ ∪ W , where W is the set of witness propositions used, Σ ∩ W = ∅. 2. Eliminate Counting: Consider, one by one, each temporal definition Ti of χf lat . Let Σi = Σ∪W ∪Xi , where Xi is a set of fresh propositions, Xi ∩Xj = ∅ for i = j. – If Ti is a temporal projection containing a C modality of the form C∼n l,u , or a UT modality of the form xUI,#b≤n y, then Lemma 6 synthesizes a formula ζi ∈ MTL over Σi such that Ti ≡ ∃Xi .ζi . – If Ti is a temporal projection containing a UT modality of the form xUI,#b≥n y, Lemma 7 gives ζi ∈ MTL over Σi such that ON FΣ (Ti ) ≡ ∃ ↓ Xi .ζi . k 3. Putting it all together : The formula ζ = i=1 ζi ∈ MTL is such that k  i=1

ON FΣ (Ti ) ≡ ∃ ↓ X.

k  i=1

ζi where X =

k 

Xi .

i=1

Lemma 6. 1. Consider a temporal definition T = w [a ↔ C≥n [l,u) b], built from Σ ∪ W . Then we synthesize a formula ζ ∈ MTL over Σ ∪ W ∪ X such that T ≡ ∃X.ζ. 2. Consider a temporal definition T = w [a ↔ xUI,#b≤n y], built from Σ ∪ W . Then we synthesize a formula ζ ∈ MTL over Σ ∪ W ∪ X such that T ≡ ∃X.ζ. Proof. 1. Lets consider intervals of the form [l, u). Our proof extends to all intervals l, u . Consider T = w [a ↔ C≥n [l,u) b]. Let ⊕ denote addition modulo n + 1.

348

S.N. Krishna et al.

Intuitively, we run a global modulo n + 1 counter B (encoded using propositional variables b0 , . . . , bn ∈ X) which is initialized to 0 and incremented modulo n + 1 at every position in timed word where b occurs, else it remains same. This is enforced by (a) and (b) below. In any interval I, there are at least n bs iff counter takes all the values of the set {0, . . . , n} in interval I. This is checked in (c) below. (a) Construction of a (Σ ∪ W, X)- simple extension. We introduce a fresh set of propositions X = {b0 , b1 , . . . , bn } and construct a simple extension ρ = (σ  , τ  ) from ρ = (σ, τ ) as follows:  – C1: σ1 = σ1 ∪ {b0 }. If bk ∈ σi and if b ∈ σi+1 , σi+1 = σi+1 ∪ {bk⊕1 }.   / σi+1 , then σi+1 = σi+1 ∪ {bk }. – C2: If bk ∈ σi and b ∈ – C3: σi has exactly one symbol from X for all 1 ≤ i ≤ |dom(ρ)|. (b) Formula specifying the above behaviour. The variables in X help in counting the number of b’s in ρ. C1 and C2 are written in MTL as follows: n n   w [(Ob ∧ bk ) → Obk⊕1 ] and δ2 = w [(O¬b ∧ bk ) → Obk ] – δ1 = k=0

k=0

(c) Marking the witness ‘a’ correctly at points satisfying C≥n [l,u) b. The index i of bi at a chosen point gives the number of b’s seen so far since the previous occurrence of b0 . From a point i, if the interval [ti + l, ti + u) has k elements of X, then there must be k b’s in [ti + l, ti + u). To mark the witness a appropriately, we need to check the number of times b occurs in [ti + l, ti + u] from the current point i. A point i ∈ dom(ρ ) is marked with witness a iff all variables of X are present in [ti + l, ti + u), as explained n  in MTL by κ = w [a ↔ ( ♦[l,u) bk )]. k=1

ζ = δ1 ∧ δ2 ∧ κ in MTL is equisatisfiable to T modulo simple projections. 2. The proof is similar to the above, details are in [6]. Lemma 7. Consider a temporal definition T = w [a ↔ xUI,#b≥n y], built from Σ ∪ W . Then we synthesize a formula ψ ∈ MTL over Σ ∪ W ∪ X such that ON FΣ (T ) ≡ ∃ ↓ X.ψ where ON FΣ (T ) is T relativized with respect to Σ. Proof. If I is of the form l, ∞), then xU l,∞),#b≥n y ≡ xU l,∞) y ∧ xU#b≥n y. The untimed threshold formula xU#b≥n y can be straightforwardly rewritten in LTL (see [7]). The case of bounded intervals is the most complex and requires oversampling. Timed word ρ is oversampled at every integer valued time point upto the maximum time stamp in ρ to give ρ . All integer points are marked by a unique proposition from C = {c0 , . . . cu } in a circular fashion with first point being marked as c0 . Each ci is associated with a bounded counter B i , which saturates at maximum value n (encoded using propositions bi0 , . . . , bin ∈ X). This counter is reset to 0 at each occurrence of ci and it is incremented each time a b occurs. This encoding is explained in (O1) and (O2) below and enforced by formula η given below. Let the resultant word after the markings be ρ . Now, by semantics, ρ , j |= xUI,#b≥n y iff for some p > j, ρ , p |= y, and x holds invariantly between j and p, and #b(ρ [j, p]) ≥ n, i.e. number of times b holds between j and p is ≥ n. This happens iff

Metric Temporal Logic with Counting

349

– either (case i) there is a nearest previous integer point to p called α (marked ci for some i) and there exist integers g, h : 0 ≤ g, h ≤ n such that #b(ρ [α, p] ≥ g , and #b(ρ [j, α, ] ≥ h and g + h ≥ n. This happens iff for some i and 0 ≤ g, h ≤ n, we have ρ , j |= xUI (y ∧ B i = g) and ρ , j |= x ∧ ¬ci U#b≥max(0,h) ci and g + h ≥ n. This is encoded by the formula δ below. – or, (case ii) j and p both lie inside a unit interval bounded by some [ci−1 , ci ]. In this case, untimed LTL formula ρ , j |= (x ∧ (¬ ∨ ci ))U#b=n y) holds. Note that the formulae are not yet relativized, the relativized ones with details are given below. (1) We consider the case when the interval I is bounded and left closed right open of the form [l, u). Our reduction below can be adapted to other kinds of bounded intervals. Let L = u−l. Define st = min(s+t, n), and s⊕t = (s+t) mod (u+1). – O1: C = {c0 , c1 , . . . , cu }. A point i of ρ is marked cg iff ti mod u = g. In the absence of such a point i (such that ti is an integer value k < t|dom(ρ)| ), we add a new point i to dom(ρ) with time stamp ti and mark it with cg iff ti mod u = g. Let ρc = (σ c , τ c ) denote the word obtained from ρ after this marking. O1 can be easily coded in MTL (say η1 ). – O2: B = ∪ui=0 B i , where B i = {bi0 , bi1 , . . . bin }. All the points of ρc marked ci are marked as bi0 . Let p, q be two integer points such that p is marked ci , q is marked ci⊕L , and no point between p, q is marked ci⊕L . p, q are L apart from c / σr+1 each other. Let p < r < q be such that big ∈ σrc for some g. If ci⊕L ∈ c i c and b ∈ σr+1 , then the point r + 1 is marked bg1 . If ci⊕L , b ∈ / σr+1 , then the point r + 1 is marked big . Each B i is a set of counters which are reset at ci and counts the number of occurrences of b up to the threshold n between a ci and the next occurrence of ci⊕L . Starting at a point marked ci with counter bi0 , the counter increments up to n on encountering a b, until the next ci⊕L . Further, we ensure that the counter B i does not appear anywhere from ci⊕L to the next ci . Let the resultant word be ρb . Let ρ be the word obtained after all the markings. (2) Formula for specifying above behaviour. We give following MTL formulae to u  (κ2i (1) ∧ κ2i (2) ∧ κ2i (3)) encodes O2 where specify O2 . η2 = i=0

w

η2i (1) =  (ci → bi0 ) ∧ η2i (2) = η2i (3) =

n  k=0 u  i=0

n  k=0

w [(O(b ∧ ¬ci⊕L ) ∧ bik ) → Obik1 ],

w [(O(¬b ∧ ¬ci⊕L ) ∧ bik ) → Obik ] and w [ci⊕L → (¬ci ∧ ¬bi )Uci ], where bi =

u  k=0

bik . Let η = η1 ∧ η2 .

(3) Marking the witness ‘a’ correctly at points satisfying xUI,#b≥n y. As explained  above there are two possible cases: Let act = (Σ ∪ W ).

350

S.N. Krishna et al.

Fig. 3. Illustration of point j, p and the point α such that tα = tj + l. α is marked with some ci since it is an integer time point. The counting of b’s is reset at ci , starting with bi0 , and continues till ci⊕L . h is the count of b’s between α and p. To satisfy (a ↔ xU[l,u),#b≥n y) at j, we check that the number of b’s between j and α is ≥ n − h when b is not true at α, and is ≥ n − h − 1 when b is true at α.

– Case(i): The formula is δ =



δ1 (g) ∧ δ2 (h) where

0≤g,h≤n∧g+h≥n

• δ1 (g) = (x ∧ ¬act)UI (y ∧ bg ) and • δ2 (h) = [{(¬act∨x)∧¬ci }U#b≥h−1 (ci ∧b)]∨[{(¬act∨x)∧¬ci }U#b≥h (ci ∧¬b)] for h > 0 and δ2 (0) = true – Case(ii): This formula takes care of the case when timestamps of points j and  p have the same integral part. Thus λ = (x ∨ ¬act) ∧ ¬c)U#b ≥n y, where c = ck captures the required condition. Note that this case is only applicable when the intervals are of the 0, u . For other type of intervals we omit this case. The formula ζ = w (a ↔ δ ∨ λ) specifies the marking of witness correctly in ρ . Thus we obtain the MTL formula ψ = η ∧ ζ. Thus, by above construction, we get a formula ψ such that ON FΣ (T ) ≡ ∃ ↓ X.ψ, where X = C ∪ B.

5

Discussion and Related Work

Within temporal and real time logics, the notion of counting has attracted considerable interest. Laroussini et al. extended untimed LTL with threshold counting constrained until operator. They showed that the expressiveness of LTL is not increased by adding threshold counting but the logic becomes exponentially more succinct. Hirshfeld and Rabinovich introduced C(0,1) operator in continuous timed QTL and showed that it added expressive power. They also showed that in continuous time, more general C l,u operator can be expressed with just C(0,1) . Building upon this, Hunter showed that MTL with C(0,1) operator is expressively complete w.r.t. FO[ 0 for some action ai for every i ≥ 0. A history is a prefix of a play. A strategy is a function π that to every history h · s assigns a probability distribution over actions such that if an action a is assigned a non-zero probability, then P (s, a) is defined. A strategy π is pure memoryless if it assigns Dirac distributions to any history and its choice depends only on the last state of the history. When we fix a strategy π, we obtain a probability measure Prπ over the set of plays. For further details, see [30]. The algorithm. It works by reduction to an MDP MG = (S  , A, P, s0 ) where – S  ⊆ S is the set of global synchronisation states; – A = C × Σ1 × Σ2 ∪ {⊥} where Σj is the set of pure memoryless strategies of player j in G that choose ∅j in every synchronisation state; – For an arbitrary state (s1 , s2 ), we define the transition function as follows: • For any (c, σ1 , σ2 ) ∈ A, the transition P ((s1 , s2 ), (c, σ1 , σ2 )) is defined if c is available in (s1 , s2 ) and the players agree in c on some action a, i.e. En(c) = {a}. If defined, the distribution P ((s1 , s2 ), (c, σ1 , σ2 )) assigns to any successor state (s1 , s2 ) the probability that in G the state (s1 , s2 ) is reached from (s1 , s2 ) via states in S \ S  by choosing c and then using the pure memoryless strategy profile (σ1 , σ2 ). • To avoid deadlocks, the transition P ((s1 , s2 ), ⊥) is defined iff no other transition is defined in (s1 , s2 ) and it is a self-loop, i.e. it assigns probability 1 to (s1 , s2 ).

366

H. Hermanns et al.

The MDP MG has size exponential in |G|. Note that all target states T are included in S  . Slightly abusing notation, let T denote the set of plays in MG that reach the set T . From standard results on MDPs [30], there exists an optimal ∗ pure memoryless strategy π ∗ , i.e. a strategy satisfying Prπ ( T ) = supπ Prπ ( T ). ∗ Furthermore, such a strategy π ∗ and the value v := Prπ ( T ) can be computed in time polynomial in |MG |. Finally, the algorithm returns TRUE if v ≥ p and FALSE, otherwise. Correctness of the algorithm Let us explain why the approach is correct. Proposition 2. The value of G is equal to the value of MG , i.e. sup Prσ ( T ) = sup Prπ ( T ). σ

π

Proof Sketch. As regards the ≤ inequality, it suffices to show that any strategy profile σ can be mimicked by some strategy π. This is simple as π in MG has always knowledge of the global state. Much more interesting is the ≥ inequality. We argue that for any strategy π there is a sequence of local strategy profiles σ 1 , σ 2 , . . . such that i lim Prσ ( T ) = Prπ ( T ). i→∞

The crucial idea is that a strategy proS2 file communicates correctly (with high prob1 2 ability) the current global state in a synchronisation state by delaying as follows. ∅1 ∅1 1 The time is divided into phases, each of ∅2 c2 (1, 1) |S1 |·2|S2 | slots (where Si are the synchroni∅1 ∅1 sation states of player i). We depict a phase S1 2 c (2, 1) ∅2 2 by the table on the right where the time c1 (3, 2) c1 (3, 1) flows from top to bottom and from left to 3 ∅2 c2 (3, 1) right (as reading a book). Players 1 and 2 try to synchronise in the row and column, sync ¬sync sync ¬sync respectively, corresponding to their current states (in circle) and in each slot take the choice ci (s1 , s2 ) optimal given the current global state is (s1 , s2 ); in the remaining slots they choose to do nothing. Since the players can change their choice only at random moments of time, their synchronising choice always stretches a bit into the successive silent slot (in a ¬sync column). The more we increase the size of each slot, the lower is the chance that a synchronisation choice of a player stretches to the next synchronisation slot. Thus, the lower is the chance of an erroneous synchronisation. We define the size of the slot to increase with i and also along the play so that for any fixed i the probability of at least one erroneous synchronisation is bounded by   κi < 1 and for i → ∞, we have κi → 0.

Distributed Synthesis in Continuous Time

6

367

Discussion and Conclusion

This paper has introduced a foundational framework for modelling and synthesising distributed controllers interacting in continuous time via handshake communication. The continuous time nature of the model induces that the interleaving scheduler has in fact very little power. We studied cooperative reachability problems for which we presented a number of undecidability results, while for non-urgent models we established decidability of both quantitative and qualitative problems for the two-player case. In the framework considered, the restriction to exponential distributions is a technical vehicle, the results could have been developed in general continuous time, e.g. by using the model of stochastic automata [9]. Distributed IMCs can be considered as an attractive base model especially in the context of information flow and other security-related studies. This is because in contrast to the discrete time setting, the power of the interleaving scheduler is no matter of debate, it can leak no essential information. From a more general perspective, distributed synthesis of control algorithms has received considerable attention in its entirety [22, 28, 29]. The asynchronous setting with handshake synchronisation has been considered in [23]. Notably, our assumption that players stay committed to a particular action choice for the time in between state changes implies the necessity to let the players explicitly solve distributed consensus problems. As done in [23], one can overcome this by letting local players pick sets of enabled actions (or letting them change choice with infinite speed), and then let some built-in magic pick a valid action from the intersection, implying that whenever possible a consensus is reached for sure. Such a change would however reintroduce the scheduler. We should point out that distributed IMCs are not fully compositional: We are assuming a fixed vector of modules, and do not discuss that modules themselves may be vectors. Otherwise, we would face the phenomenon of autoconcurrency [10], where transitions with identical synchronisation actions might get enabled concurrently, despite not synchronising. This in turn would again re-introduce distinguishing power of the scheduler. Distributed Markov chains [31] constitute another recent discrete-time approach where interleaving nondeterminism is tamed successfully via assumptions on the communication structure. The observation that continuous time reduces the power of the interleaving scheduler is not entirely new. Though not explicitly discussed, it underpins the model of probabilistic I/O automata (PIOA) [34] which uses I/O communication with input-enabledness and output-determinism. In that setting, output-determinism implies that local players have no decisive power, and hence a continuous time Markov chain arises. We can approximate I/O-based communication by distributed IMCs without the need for outputdeterminism. The approximation is linked to arbitrarily small but non-zero delays needed to cycle through synchronising action sets. A profound investigation of the continuous-time particularities of this and other synchronisation disciplines is considered an interesting topic for future work.

368

H. Hermanns et al.

Acknowledgements. This work is supported by the EU 7th Framework Programme projects 295261(MEALS) and 318490 (SENSATION), by the Czech Science Foundation project P202/12/G061, the DFG Transregional Collaborative Research Centre SFB/TR 14 AVACS, and by the CDZ project 1023 (CAP).

References 1. Baier, C., Gr¨ oßer, M., Bertrand, N.: Probabilistic ω-automata. J. ACM 59(1), 1 (2012) 2. Baier, C., Katoen, J.P.: Principles of Model Checking. MIT Press, Cambridge (2008) 3. Bernstein, D.S., Givan, R., Immerman, N., Zilberstein, S.: The complexity of decentralized control of Markov decision processes. Math. Oper. Res. 27(4), 819–840 (2002) 4. Bertoni, A., Mauri, G., Torelli, M.: Some recursively unsolvable problems relating to isolated cutpoints in probabilistic automata. In: Degano, P., Gorrieri, R., Marchetti-Spaccamela, A. (eds.) ICALP 1997. LNCS, vol. 1256. Springer, Heidelberg (1997) 5. Berwanger, D., Kaiser, L.: Information tracking in games on graphs. J. Logic Lang. Inform. 19(4), 395–412 (2010) 6. Brookes, S.D., Hoare, C.A.R., Roscoe, A.W.: A theory of communicating sequential processes. J. ACM 31(3), 560–599 (1984) 7. Canetti, R., Cheung, L., Kaynar, D.K., Liskov, M., Lynch, N.A., Pereira, O., Segala, R.: Analyzing security protocols using time-bounded task-PIOAs. Discrete Event Dyn. Syst. 18(1), 111–159 (2008) 8. Cheung, L., Lynch, N.A., Segala, R., Vaandrager, F.W.: Switched PIOA: parallel composition via distributed scheduling. Theoret. Comput. Sci. 365(1–2), 83–108 (2006) 9. D’Argenio, P.R., Katoen, J.-P.: A theory of stochastic systems part I: stochastic automata. Inf. Comput. 203(1), 1–38 (2005) 10. Droste, M., Gastin, P.: Asynchronous cellular automata for pomsets without autoconcurrency. In: Sassone, V., Montanari, U. (eds.) CONCUR 1996. LNCS, vol. 1119. Springer, Heidelberg (1996) 11. Genest, B., Gimbert, H., Muscholl, A., Walukiewicz, I.: Asynchronous games over tree architectures. In: Fomin, F.V., Freivalds, R., Kwiatkowska, M., Peleg, D. (eds.) ICALP 2013, Part II. LNCS, vol. 7966, pp. 275–286. Springer, Heidelberg (2013) 12. Gimbert, H., Oualhadj, Y.: Probabilistic automata on finite words: decidable and undecidable problems. In: Abramsky, S., Gavoille, C., Kirchner, C., Meyer auf der Heide, F., Spirakis, P.G. (eds.) ICALP 2010. LNCS, vol. 6199, pp. 527–538. Springer, Heidelberg (2010) 13. Giro, S., D’Argenio, P.R., Fioriti, L.M.F.: Distributed probabilistic input/output automata: Expressiveness, (un)decidability and algorithms. Theoret. Comput. Sci. 538, 84–102 (2014) 14. Guck, D., Han, T., Katoen, J.-P., Neuh¨ außer, M.R.: Quantitative timed analysis of interactive markov chains. In: Goodloe, A.E., Person, S. (eds.) NFM 2012. LNCS, vol. 7226, pp. 8–23. Springer, Heidelberg (2012) 15. Hermanns, H., Johr, S.: May we reach it? Or must we? In what time? With what probability? In: MMB, pp. 125–140. VDE Verlag (2008) 16. Hermanns, H.: Interactive Markov Chains and The Quest for Quantified Quality. Lecture Notes in Computer Science, vol. 2428. Springer, New York (2002)

Distributed Synthesis in Continuous Time

369

17. Hermanns, H., Vester, S., Krˇca ´l, J.: Distributed synthesis in continuous time. CoRR abs/1601.01587 (2016) 18. Hermanns, H., Katoen, J.-P.: The how and why of interactive markov chains. In: Boer, F.S., Bonsangue, M.M., Hallerstede, S., Leuschel, M. (eds.) FMCO 2009. LNCS, vol. 6286, pp. 311–338. Springer, Heidelberg (2010) 19. Hermanns, H., Krˇc´ al, J., Kˇret´ınsk´ y, J.: Compositional verification and optimization of interactive markov chains. In: D’Argenio, P.R., Melgratti, H. (eds.) CONCUR 2013 – Concurrency Theory. LNCS, vol. 8052, pp. 364–379. Springer, Heidelberg (2013) 20. Katoen, J.-P., Klink, D., Neuh¨ außer, M.R.: Compositional abstraction for stochastic systems. In: Ouaknine, J., Vaandrager, F.W. (eds.) FORMATS 2009. LNCS, vol. 5813, pp. 195–211. Springer, Heidelberg (2009) 21. Katoen, J.-P., Zapreev, I.S., Hahn, E.M., Hermanns, H., Jansen, D.N.: The ins and outs of the probabilistic model checker MRMC. Perform. Eval. 68(2), 90–104 (2011) 22. Madhusudan, P., Thiagarajan, P.S.: Distributed controller synthesis for local specifications. In: Orejas, F., Spirakis, P.G., van Leeuwen, J. (eds.) ICALP 2001. LNCS, vol. 2076, pp. 396–407. Springer, Heidelberg (2001) 23. Madhusudan, P., Thiagarajan, P.S.: A decidable class of asynchronous distributed controllers. In: Brim, L., Janˇcar, P., Kˇret´ınsk´ y, M., Kuˇcera, A. (eds.) CONCUR 2002. LNCS, vol. 2421, pp. 145–160. Springer, Heidelberg (2002) 24. Milner, R.: Calculi for synchrony and asynchrony. Theoret. Comput. Sci. 25, 267– 310 (1983) 25. Neuts, M.F.: Matrix-Geometric Solutions in Stochastic Models: An Algorithmic Approach. Courier Corporation, New York (1981) 26. Paz, A.: Introduction to Probabilistic Automata. Academic Press Inc., London (1971) 27. Pelozo, S.S., D’Argenio, P.R.: Security analysis in probabilistic distributed protocols via bounded reachability. In: Palamidessi, C., Ryan, M.D. (eds.) TGC 2012. LNCS, vol. 8191, pp. 182–197. Springer, Heidelberg (2013) 28. Pnueli, A., Rosner, R.: On the synthesis of an asynchronous reactive module. In: Ausiello, G., Dezani-Ciancaglini, M., Della Rocca, S.R. (eds.) ICALP. LNCS, vol. 372, pp. 652–671. Springer, Heidelberg (1989) 29. Pnueli, A., Rosner, R.: Distributed reactive systems are hard to synthesize. In: FOCS, pp. 746–757. IEEE Computer Society (1990) 30. Puterman, M.L.: Markov Decision Processes: Discrete Stochastic Dynamic Programming. John Wiley & Sons, New York (2009) 31. Saha, R., Esparza, J., Jha, S.K., Mukund, M., Thiagarajan, P.S.: Distributed markov chains. In: D’Souza, D., Lal, A., Larsen, K.G. (eds.) VMCAI 2015. LNCS, vol. 8931, pp. 117–134. Springer, Heidelberg (2015) 32. Segala, R.: Modeling and Verification of Randomized Distributed Real-Time Systems. Ph.D. thesis, Massachusetts Institute of Technology (1995) 33. Sinopoli, B., Sharp, C., Schenato, L., Schaffert, S., Sastry, S.S.: Distributed control applications within sensor networks. Proc. IEEE 91, 1235–1246 (2003) 34. Sue-Hwey, W., Smolka, S.A., Stark, E.W.: Composition and behaviors of probabilistic I/O automata. Theoret. Comput. Sci. 176(1–2), 1–38 (1997)

Proof Theory and Lambda Calculus

Unary Resolution: Characterizing PTIME Cl´ement Aubert1 , Marc Bagnol2 , and Thomas Seiller3(B) 1

2

Department of Computer Science, Appalachian State University, Boone, USA Department of Mathematics and Statistics, University of Ottawa, Ottawa, Canada 3 Department of Computer Science, University of Copenhagen, Copenhagen, Denmark [email protected]

Abstract. We give a characterization of deterministic polynomial time computation based on an algebraic structure called the resolution semiring, whose elements can be understood as logic programs or sets of rewriting rules over first-order terms. This construction stems from an interactive interpretation of the cut-elimination procedure of linear logic known as the geometry of interaction . This framework is restricted to terms (logic programs, rewriting rules) using only unary symbols, and this restriction is shown to be complete for polynomial time computation by encoding pushdown automata. Soundness w.r.t. Ptime is proven thanks to a saturation method similar to the one used for pushdown systems and inspired by the memoization technique. A Ptime-completeness result for a class of logic programming queries that uses only unary function symbols comes as a direct consequence. Keywords: Implicit complexity · Unary queries · Logic programming · Geometry of interaction · Proof theory · Pushdown automata · Saturation · Memoization

1

Introduction

Complexity theory classifies computational problems relatively to the amount of time or memory needed to solve them. Complexity classes are defined as sets of problems that can be solved by algorithms whose executions need comparable amounts of resources. For instance, the class Ptime is the set of predicates over binary words that can be decided by a Turing machine whose execution time is bounded by a polynomial in the size of its input. One of the main motivations for an implicit computational complexity (ICC) theory is to find machine-independent characterizations of complexity classes. The aim is to characterize classes not “by constraining the amount of resources a machine is allowed to use, but rather by imposing linguistic constraints on the This work was partly supported by the ANR-14-CE25-0005 Elica, the ANR-11-INSE0007 Rever, the ANR-10-BLAN-0213 Logoi, the ANR-11-BS02-0010 Recr´e, the ANR 12 JS02 006 01 project Coquas and the European Union’s Marie Sklodowska-Curie Individual Fellowship (H2020-MSCA-IF-2014) 659920 - ReACT. c Springer-Verlag Berlin Heidelberg 2016  B. Jacobs and C. L¨ oding (Eds.): FOSSACS 2016, LNCS 9634, pp. 373–389, 2016. DOI: 10.1007/978-3-662-49630-5 22

374

C. Aubert et al.

way algorithms are formulated.” [18, p. 90] This has been already achieved via different approaches, for instance by considering restricted programming languages or computational principles [12,37, 38]. A number of results in this area also arose from proof theory, through the study of subsystems of linear logic [25]. More precisely, the Curry-Howard—or proofs as programs—correspondence expresses an isomorphism between formal proofs and typed programs. In this approach, once a formula Nat corresponding to the type of binary integers is set, proofs of the formula Nat ⇒ Nat are algorithms computing functions from integers to integers, via the cut-elimination procedure. By considering restricted subsystems, one allows less proofs of type Nat ⇒ Nat, hence less algorithms can be implemented, and the class of accepted proofs, or programs, may correspond1 to some complexity class: elementary complexity [21, 29], polynomial time [11, 36], logarithmic [19] and polynomial [24] space. More recently, new methods for obtaining implicit characterizations of complexity classes based on the geometry of interaction (GoI) research program [27] have been developed. The GoI approach offers a more abstract and algebraic point of view on the cut-elimination procedure of linear logic. One works with a set of untyped programs represented as some geometric objects, e.g. graphs [20, 40] or generalizations of graphs [42], bounded linear maps between Hilbert spaces (operators) [26, 30,41], clauses (or “flows”) [8,28]. This set of objects is then considered together with an abstract notion of execution, seen as an interactive procedure: a function does not process a static input, but rather communicate with it, asking for values, reading its answers, asking for another value, etc. (this interactive point of view on computation has proven crucial in characterizing logarithmic space computation [19]). This method does not proceed by restricting a type system, but by imposing original conditions, of an algebraic nature, on the representation of programs. Note that one still benefits from the work in the typed case: for instance, the representation of words used here directly comes from their representation in linear logic. The first results in this direction were based on operator algebra [5, 6,31]. This paper considers a more syntactic flavor of the GoI interpretation, where untyped programs are represented in the so-called resolution semiring [8], a semiring based on the resolution rule [39] and a specific class of logic programs. This setting presents some advantages: it avoids the heavy structure of operator algebras theory, eases the discussions in terms of complexity (as first-order terms have natural notions of size, height, etc.) and offers a straightforward connection with complexity of logic programming [22]. Previous works in this direction led to characterizations of logarithmic space predicates Logspace and co-NLogspace [2,3], by considering restrictions on the height of variables. The main contribution of this paper is a characterization of the class Ptime by studying a natural limitation, the restriction to unary function symbols. Pushdown automata are easily related to this simple restriction, for they can be represented as logical programs satisfying this “unarity” restriction. This implies the completeness of the model under consideration for polynomial time 1

In an extensional correspondence: they can compute the same functions.

Unary Resolution: Characterizing PTIME

375

predicates. Soundness follows from a variation of the saturation algorithm for pushdown systems [13], inspired by S. Cook’s memoization technique [17] for pushdown automata, that proves that any such unary logic program can be decided in polynomial time. Compared to other ICC characterizations of Ptime, and in particular those coming from proof theory, this method has a simple formulation and provides an original point of view on this complexity class. It also constitutes the first characterization of a time-complexity class directly on the semantic side of the GoI interpretation. Nevertheless, the results presented here can be read independently of any knowledge of GoI. An immediate consequence of this work is a Ptime-completeness result for a specific class of logic programming queries corresponding to unary flows. 1.1

Outline

Section 2.1 gives the formal definition of the resolution semiring; then representation of words and programs in this structure is briefly explained (Sect. 2.2). Section 2.3 introduce the restricted semiring that will be under study, the Stack semiring. We believe the technical results presented in this section to be of importance, as they describe an algebraic restriction corresponding to Ptime and broaden previous algebraic restrictions for (co-N)Logspace. The next two sections are respectively devoted to the completeness and soundness results for Ptime. Proving completeness needs to first review multihead finite automata with pushdown stack, that characterize Ptime, and then to represent them as elements built from the Stack semiring (Sect. 3). The soundness result is then obtained by “saturating” elements of the stack semiring, so that they become decidable with Ptime resources (Sect. 4). Ptime-completeness of unary logic programming queries is then proved to be implied by this result (Sect. 5). Sketched proofs are detailed in a technical report [4] and in the second author’s Ph.D. thesis [8].

2 2.1

The Resolution Semiring Flows and Wirings

We begin with some reminders on first-order terms and unification theory. Notation 1 (terms). We consider first-order terms, written t, u, v, . . ., built from variables and function symbols with assigned finite arity. Symbols of arity 0 will be called constants. Sets of variables and of function symbols of any arity are supposed infinite. Variables will be noted in italics font (e.g. x, y) and function symbols in typewriter font (e.g. c, f(·), g(·, ·)). We distinguish a binary function symbol • (in infix notation) and a constant symbol . We will omit the parentheses for • and write t • u • v for t • (u • v).

376

C. Aubert et al.

We write var(t) the set of variables occurring in the term t and say that t is closed if var(t) = ∅. The height h (t) of a term t is the maximal distance between its root and leaves; a variable occurrence’s height in t is its distance to the root. We will write θt for the result of applying the substitution θ to the term t and will call renaming a substitution α that bijectively maps variables to variables. We focus on the resolution of equalities t = u between terms, and hence need some definitions. Definition 2 (Unification, Matching and Disjointness). Two terms t, u are: – unifiable if there exists a substitution θ—a unifier of t and u—such that θt = θu. If any other unifier of t and u is such that there exists a substitution that maps it to θ, we say θ is the most general unifier (MGU) of t and u; – matchable if t , u are unifiable, where t , u are renamings of t, u such that var(t ) ∩ var(u ) = ∅; – disjoint if they are not matchable. A fundamental result of unification theory is that when two terms are unifiable, a MGU exists and is computable. More specifically, the problem of deciding whether two terms are unifiable is Ptime-complete [23, Theorem 1]. The notion of MGU allows to formulate the resolution rule, a key concept of logic programming defining the composition of Horn clauses (expressions of the form H  B1 , . . . , Bn ):

Note that the condition on variables implies that we are matching U and V rather than unifying them. In other words, the resolution rule deals with variables as if they were bounded. From this perspective, “flows”—defined below—are a specific type of Horn clauses H  B, with exactly one formula B at the right of  and all the variables of H already occurring in B. The product of flows will be defined as the resolution rule restricted to this specific type of clauses. Definition 3 (Flow). A flow is an ordered pair f of terms f := t  u, with var(t) ⊆ var(u). Flows are considered up to renaming: for any renaming α, t  u = αt  αu. A flow can be understood as a rewriting rule over the set of first-order terms, acting at the root. For instance, the flow g(x)  f(x) corresponds to “rewrite terms of the form f(v) as g(v)”. Next comes the definition of the product of flows. From the term-rewriting perspective, this operation corresponds to composing two rules—when possible, i.e. when the result of the first rewriting rule allows the application of the

Unary Resolution: Characterizing PTIME

377

second—into a single one. For instance, one can compose the flows f1 := h(x)  g(x) and f2 := g(f(x))  f(x) to produce the flow f1 f2 = h(f(x))  f(x). Notice by the way that this (partial) product is not commutative, as composing these rules the other way around is impossible, i.e. f2 f1 is not defined. Definition 4 (Product of Flows). Let t  u and v  w be two flows. Suppose we picked representatives of the renaming classes such that var(u) ∩ var(v) = ∅. The product of t  u and v  w is defined when u and v are unifiable, with MGU θ, as (t  u)(v  w) := θt  θw. We now define wirings, which are just finite sets of flows and therefore correspond to logic programs. From the term-rewriting perspective they are just sets of rewriting rules. The definition of product of flows naturally lifts to wirings. Definition 5 (Wiring). A wiring is a finite set of flows. Their product is defined as F G := {f g|f ∈ F, g ∈ G, f g defined}. The resolution semiring R is the set of all wirings. The set of wirings R indeed enjoys a structure of semiring2 . We will use an additive notation for sets of flows to highlight this situation: – The symbol + will be used in place of ∪, and we write sets as sums of their elements: {f1 , . . . , fn } := f1 + · · · + fn . – We denote by 0 the empty set, i.e. the unit of the sum. – The unit for the product is the wiring I := x  x. As we will always be working within R, the term “semiring” will be used instead of “subsemiring of R”. Finally, let us extend the notion of height to flows and wirings, and recall the definition of nilpotency. Definition 6 (Height). The height h(f ) of a flow f = t  u is defined as max{h(t), h(u)}. A wiring’s height is defined as h(F ) = max{h(f )|f ∈ F }. By convention h(0) = 0. Definition 7 (Nilpotency). A wiring F is nilpotent—written Nil(F )—if and only if F n = 0 for some n. That standard notion of nilpotency, coming from abstract algebra, has a specific reading in our case. In terms of logic programming, it means that all chains obtained by applying the resolution rule to the set of clauses we consider cannot be longer than a certain bound. From a rewriting point of view, it means that the set of rewriting rules is terminating with a uniform bound on the length of rewriting chains —again, we consider rewriting at the root of terms, while general term rewriting systems [7] allow for in-context rewriting. 2

A semiring is a set R equipped with two operations + (the sum) and × (the product), and an element 0 ∈ R such that: (R, +, 0) is a commutative monoid; (R, ×) is a semigroup; the product distributes over the sum; and the element 0 is absorbent, i.e. 0 × r = r × 0 = 0 for all r ∈ R.

378

2.2

C. Aubert et al.

Representation of Words and Programs

This section explains and motivates the representation of words as flows. Studying their interaction with wirings from a specific semiring allows to define notions of program and accepted language. First, let us extend the binary function symbol • , used to construct terms, to flows and then semirings. Definition 8. Let u  v and t  w be two flows. Suppose we have chosen representatives of their renaming classes that have disjoint sets of variables. We define • (t  w) := u • t  v • w. The operation is extended to  (u  v) wirings by ( i fi ) • ( j gj ) :=  i,j fi • gj . Then, given two semirings A and B, we define the semiring A • B := { i Fi • Gi |Fi ∈ A , Gi ∈ B}. The operation indeed defines a semiring because for any wirings F , F  , G and G , we have (F • G)(F • G) = F F  • GG . Moreover, we carry on the convention of writing A • B • C for A • (B • C). 

Notation 9. We write t  u the sum t  u + u  t. Definition 10 (Word Representation). We fix for the rest of this paper an infinite set of constant symbols P (the position constants) and a finite alphabet Σ disjoint from P with  ∈ Σ (we write Σ ∗ the set of words over Σ). Let W = c1 · · · cn ∈ Σ ∗ and p = p0 , p1 , . . . , pn be pairwise distinct elements of P. Writing pn+1 = p0 and cn+1 = c0 = , we define the representation of W associated with p0 , p1 , . . . , pn as the following wiring, where x and y are two arbitrary but distinct variables: ¯p = W

n 

ci • r • x • y • head(pi )  ci+1 • l • x • y • head(pi+1 )

i=0

Position constants p0 , p1 , . . . , pn represent memory cells storing the symbols , c1 , c2 , . . . . The representation of words is dynamic, i.e. we may think intuitively of movement instructions for an automaton reading the input, getting instructions to move from a symbol to the next (at the left, hence the l) or the previous (to the right, hence the r). More details on this point of view will be given in the proof of Theorem 4. A term ci • r (resp. ci • l) at position p will be linked by flows of the representation to an element ci+1 • l at position pi+1 (resp. ci−1 • r at position pi−1 ). Taking cn+1 = c0 =  reflects the Church encoding of words and make the representation of the input circular. Flows representing the word c1 · · · cn can be pictured as follows:

The notion of observation will be the counterpart of a program in our construction. We first give a general definition, that will be instantiated later to

Unary Resolution: Characterizing PTIME

379

particular classes of observations characterizing complexity classes. The important point is that observations cannot use any position constant, so that they ¯ p of a word W . interact the same way with all the representations W Definition 11 (Observation Semiring). We define the semirings P⊥ of flows that do not use the symbols in P; and Σlr the semiring generated by flows of the form c • d  c • d with c, c ∈ Σ ∪ {} and d, d ∈ {l, r}. We define the semiring of observations as O := (Σlr • R) ∩ P⊥ , and the semiring of observations over the semiring A as O[ A ] := (Σlr • A) ∩ P⊥ . The following property is a consequence of the fact that observations cannot use position constants [8, Theorem IV.5]. ¯ q be two representations of a word ¯ p and W Theorem 1 (normativity). Let W ¯ q ). ¯ p ) if and only if Nil(OW W and O an observation. Then Nil(OW How a word can be accepted by an observation can now be safely defined: the following definition is independent of the specific choice of a representation. Definition 12 (Accepted Language). Let O be an observation. The lan¯ p )}. guage accepted by O is defined as L(O) := {W ∈ Σ ∗ | ∀p, Nil(OW A previous work [3] investigated the semiring of balanced wirings, that are defined as sets of balanced—or “height-preserving”—flows. Definition 13 (Balance). A flow f = t  u is balanced if for any variable x ∈ var(t) ∪ var(u), all occurrences of x in both t and u have the same height (recall notations p. 4). A balanced wiring F is a sum of balanced flows and the set of balanced wirings is written Rb . A balanced observation is an element of O[ Rb • Rb ]. This simple restriction was shown to characterize (non-deterministic) logarithmic space computation [3, Theorems 34-35], with a natural subclass of balanced wirings corresponding to the deterministic case. The balanced restriction won’t be further considered, even if previous results on the nilpotency problem for balanced wirings [3, p. 54], [8, Theorem IV.12] are required to complete the detailed proof of Theorem 5 [4, 8]. 2.3

The Stack Semiring

This paper deals with another restriction on flows, namely the restriction to unary flows, i.e. defined with unary function symbols only. The semiring of wirings composed only of unary flows is called the Stack semiring, and will be shown to give a characterization of polynomial time computation. Below are the needed definitions and results about this semiring, a more complete picture is in the second author’s Ph.D. thesis [8]. Definition 14 (Unary Flows). A unary flow is a flow built using only unary function  symbols and a variable. The semiring Stack is the set of wirings of the form i ti  ui where the ti  ui are unary flows.

380

C. Aubert et al.

Example 1 The flows f(f(x ))  g(x ) and x  g(x ) are both unary, while x • f(x)  g(x) and f(c)  x are not. The notion of cyclic flow is crucial to prove the characterization of polynomial time computation. It is complementary to the nilpotency property for elements of Stack , i.e. a wiring in Stack will be either cyclic or nilpotent. Definition 15 (Cyclicity). A flow t  u is a cycle if t and u are matchable (Definition 2). A wiring F is cyclic if there is a k such that F k contains a cycle. Remark 1. A flow f is a cycle iff f 2 = 0, which in turn implies f n = 0 for all n in the case f is unary. This does not hold in general: f = x • c  d • x is a cycle as f 2 = c • c  d • d = 0, but f 3 = (x • c  d • x )(c • c  d • d) = 0. Theorem 2 (nilpotency). A wiring F ∈ Stack is nilpotent iff it is acyclic. Proof (Sketch [8, Theorem II.52]). An immediate consequence of Remark 1 is that if F is acyclic, then it is not nilpotent. The converse is a consequence of a bound on the height of elements of F n when F is acyclic [10]. From this, a contradiction can be obtained by realizing that manipulating bounded height terms built from a finite pool of symbols implies that one is wandering in a finite set and will eventually be cycling in it. Example 2. The following nilpotent element of Stack illustrates how the nilpotency problem can be tricky to solve efficiently: F :=

f1 (x)  f0 (x) + f0 (f1 (x))  f1 (f0 (x)) + f0 (f0 (f1 (x)))  f1 (f1 (f0 (x))) + f0 (f0 (f0 (x)))  f1 (f1 (f1 (x)))

Taking the sequence fx fy fz to be the integer x + 2y + 4z, this wiring implements a counter from 0 to 7 in binary notation, that resets to 0 when it reaches 8. It is clear with this intuition in mind that this wiring is cyclic. Indeed, an easy computation shows that f0 (f0 (f0 (x)))  f0 (f0 (f0 (x))) ∈ F 8 . Lifting this example to the case of a counter from 0 to 2n − 1, gives a wiring for which the number of iterations needed to find a cycle is exponential in its size. This rules out a polynomial time decision procedure for the nilpotency problem that would simply compute iterations of a wiring until it finds a cycle. Finally, let us define a new class of observations, based on the Stack semiring. Definition 16 (Balanced Observation with Stack). A balanced observation with stack is an element of the semiring Ob+s := O[ Stack • Rb ].

Unary Resolution: Characterizing PTIME

3

381

Pushdown Automata and PTIME Completeness

Automata form a simple model of computation that can be extended in different ways. For instance, allowing multiple heads that can move in two directions on the input tape gives a model of computation equivalent to read-only Turing machines. If one adds moreover a “pushdown stack” one defines “pushdown automata”, well-known to capture polynomial-time computation. Ptimecompleteness of balanced observation with stacks will be attained by encoding pushdown automata: we recall briefly their definition and characterization of Ptime, before sketching how to represent them as observations. Definition 17 (Pushdown Automata (2MFA+S)). For k  1, a pushdown automaton (formally, a 2-way k-head finite automaton with pushdown stack (2MFA+S(k))) is a tuple M = {S, i, A, B, , , , σ} where: – – – – –

S is the finite set of states, with i ∈ S the initial state; A is the input alphabet, B the stack alphabet;  and  are the left and right endmarkers, , ∈ / A;  is the bottom symbol of the stack,  ∈ / B; σ is the transition relation, i.e. a subset of the product (S×(A )k ×B )×(S× {−1, 0, +1}k ×{pop, push(b)}) where A (resp. B ) denotes A∪{, } (resp. B ∪ {}). The instruction −1 corresponds to moving the head one cell to the left, 0 corresponds to keeping the head on the current cell and +1 corresponds to moving it one cell to the right. Regarding the pushdown stack, the instruction pop means “erase the top symbol”, while, for all b ∈ B, push(b) means “write b on top of the stack”.

The automaton rejects the input if it loops, otherwise it accepts. This condition is equivalent to the standard way of defining acceptance and rejection by “reaching a final state” [35, Theorem 2]. Modulo another standard transformation, we restrict the transition relation so that at most one head moves at each transition. We used in our previous work [3,6] the characterization of Logspace and NLogspace by 2-way k-head finite automata without pushdown stacks [43, pp. 223–225]. The addition of a pushdown stack improves the expressiveness of the machine model, as stated in the following theorem. Theorem 3. Pushdown automata characterize Ptime. Proof. Without reproving this classical result of complexity theory, we review the main ideas supporting it. Simulating a Ptime Turing machine with a Pushdown automata amounts to designing an equivalent Turing machine whose movements of heads follow a regular pattern. That permits to seamlessly simulate their contents with a pushdown stack. A complete proof [16, pp. 9–11] as well as a precise algorithm [43, pp. 238–240] can be found in the literature.

382

C. Aubert et al.

Simulating a Pushdown automata with a Polynomial-time Turing Machine. cannot amount to simply simulate step-by-step the automaton with the Turing machine. The reason is that for any pushdown automaton, one can design a pushdown automaton that recognizes the same language but runs exponentially slower [1, p. 197]. That the pushdown automaton can accept its input after an exponential computation time is similar with the situation of the counter in Exmaple 2. The technique invented by Alfred V. Aho et al. [1] and made popular by Stephen A. Cook consists in building a “memoization table” allowing the Turing machine to create shortcuts in the simulation of the pushdown automaton, decreasing drastically its computation time. In some cases, an automaton with an exponentially long run can even be simulated in linear time [17]. Let us now consider the proof of Ptime-completeness for the set of balanced observations with stacks. It relies on an encoding that is similar to the previously developed encoding of 2-way k-head finite automata (without pushdown stack) by flows [3, Sect. 4.1]. The only difference is the addition of a “plug-in” that allows to represent stacks in observations. Remember that acceptance by observations is phrased in terms of nilpo¯ p of the observation and the representation of the input tency of the product OW (Definition 12). Hence the computation in this model is defined as an itera¯ p )2 , (OW ¯ p )3 , . . . and ¯ p , (OW tion: one computes by considering the sequence OW ¯ p )n = 0 the computation either ends at some point (i.e. accepts)—that is (OW for some integer n—or loops (i.e. rejects). This iteration represents a dialogue between the observation and its input: whereas an automaton is often thought of as manipulating some the “passive” data, in our setting, the observation and the word representation interact, taking turns in making the situation evolve. Theorem 4. If L ∈ Ptime, then there exists a balanced observation with stack O ∈ Ob+s such that L = L(O). Proof. Let A = Σ be the input alphabet and M the 2MFA+S(k + 1) that recognizes L. By Theorem 3, such a M exists, and its transition relation is encoded as a balanced observation with stack (Definition 16). More precisely, the automaton will be represented as an element OM of Ob+s = O[ Stack • Rb ] which can be written as a sum of flows of the form c • d • σ(x) • q • auxk (y1 , . . . , yk ) • head(z  )  c • d • s(x) • q • auxk (y1 , . . . , yk ) • head(z) with – – – – –

c, c ∈ Σ ∪ {} and d, d ∈ {l, r}, σ a finite sequence of unary function symbols, s a unary function symbol, q, q two constant symbols, auxk and head two functions symbols of respective arity k and 1.

Unary Resolution: Characterizing PTIME

383

The intuition behind the encoding is that a configuration of a 2MFA+S(k + 1) processing an input can be seen as a closed term c • d • τ () • q • auxk (pi1 , . . . , pik ) • head(pj ) where the pi are position constants representing the positions of the main pointer (head(pj )) and of the auxiliary pointers (auxk (pi1 , . . . , pik )); the symbol q represents the state the automaton is in; τ () represents the current stack; the symbol d represents the direction of the next move of the main pointer; the symbol c represents the symbol currently read by the main pointer. When a configuration matches the right side of the flow, the transition is followed, leading to an updated configuration. ¯ p , the product of the encoding of M More precisely, the iterations of OM W with a word representation, is observed. Let us now explain how the basic operations of M are simulated: Moving the Pointers. Looking back at the definition of the encoding of words (Definition 10) gives a new reading of the action of the representation of a word: it moves the main pointer in the required direction. From that perspective, the position holding the symbol  in Definition 10 allows to simulate the behavior of the endmarkers  and . On the other hand, the observation is not able to manipulate the position of pointers directly (remember observations are forbidden to use the position constants) but can change the direction symbol d, rearrange pointers (hence changing which one is the main pointer) and modify its state and the symbol c accordingly. For instance, a flow of the form · · · • auxk (x, . . . , yk ) • head(y1 )  · · · • auxk (y1 , . . . , yk ) • head(x) encodes the instruction “swap the main pointer and the first auxiliary pointer”. Note however that our model has no built-in way to remember the values of the auxiliary pointers—it remembers only their positions as arguments of auxk (· · · )—, but this can be implemented easily using additional states. Handling the Stack. Given a unary function symbol b(·) for each symbol b of the stack alphabet B , reading the stack, pushing and popping elements are easily implemented: · · · • x • · · ·  · · · • b(x) • · · · · · · • c(b(x)) • · · ·  · · · • b(x) • · · ·

(Read b and pop it) (Read b and push c)

Changing the State. Given a constant q for each state q of M , updating the state amounts to picking the right q and q in the flow representing the transition. Acceptance and Rejection. The encoding of acceptance and rejection is slightly more delicate, as detailed in a previous article [5, Sect. 6.2.3.].

384

C. Aubert et al.

The basic idea is that acceptance in our model is defined as nilpotency, that is to say: the absence of loops. If no transition in the automaton can be fired, then no flow in our encoding can be unified, and the computation ends. Conversely, a loop in the automaton will refrain the wiring from being nilpotent. Loops should be represented as a re-initialization of the computation, so that the observation performs forever the same computation when rejecting the input. Another encoding may interfere with the representation of acceptation as termination: an “in-place loop” triggered when reaching a particular state would make the observation cyclic, hence preventing the observation from being nilpotent no matter the word representation processed. Indeed, the “loop” in Definition 17 of pushdown automata is to be read as “perform forever the same computation”. Observations resulting from encoding pushdown automata are sums of flows of a particular form (shown at the beginning of the preceding proof). However, using general observations with stack, not constrained in this way, does not increase the expressive power: the next section is devoted to prove that the language recognized by any observation with stack lies in Ptime.

4

Nilpotency in Stack and PTIME Soundness

We now introduce the saturation technique, which allows to decide nilpotency of Stack elements in polynomial time. This technique relies on the fact that in certain cases, the height of flows does not grow when computing their product. It adapts memoization [32] to our setting: we repeatedly extend the wiring by adding pairwise products of flows, allowing for more and more “transitions”. Remark 2. As pointed out by a reviewer of a previous version of this work, deciding the nilpotency of a unary flow is reminiscent of the problem of acyclicity for the configuration graph of a pushdown system (PDS) [13], a problem known to lie in Ptime [15]. However, our algorithm treats every state of the corresponding PDS as initial, and would detect cycles even in non-connected components: our problem is probably closer to the “uniform halting problem” [34], a problem known to be decidable [14, p. 10]. Whether this last problem, equivalent to deciding the nœth´eriennit´e of a finite system rewriting suffix words, is known to lie in Ptime, and if our Theorem 5 entails that bound, are both unknown to us.   Notation of unary function symbols. If h τ (x) ≥   18. Let  τ and  σ be  sequences  h σ(x) (reps. h τ (x) ≤ h σ(x) ), we say that τ (x)  σ(x) is increasing (resp. decreasing). A wiring in Stack is increasing (resp. decreasing) if it contains only increasing (resp. decreasing) unary flows. Lemma 1 (stability of height). Let τ and σ be sequences of unary function symbols. If τ is decreasing and σ is increasing, then h(τ σ) ≤ max{h(τ ), h(σ)}.

Unary Resolution: Characterizing PTIME

385

With this lemma in mind, we can define a shortcut operation that augments an element of Stack by adding new flows while keeping the maximal height unchanged. Iterating this operation, we obtain a saturated version of the initial wiring, containing shortcuts, shortcuts of shortcuts, etc. In a sense we are designing an exponentiation by squaring procedure for elements of Stack , the algebraic reading of memoization in our context. Definition 19 (Saturation). If F ∈ Stack we define its increasing F ↑ := {f ∈ F |f is increasing} and decreasing F ↓ := {f ∈ F |f is decreasing} subsets. We set the shortcut operation short(F ) := F + F ↓ F ↑ and its least fixpoint, which we call the saturation of F : satur(F ) := n∈N shortn (F ) (where shortn denotes the nth iteration of short). The point of this operation is that it is computable in Ptime (the fixpoint is reached in polynomial time) because of Lemma 1. This leads to a Ptime decision procedure for nilpotency of elements of Stack . Theorem 5 (nilpotency is in PTIME). Given any integer h, there is a procedure Nilph (·) ∈ Ptime that, given a F ∈ Stack such that h(F ) ≤ h as an input, accepts iff F is nilpotent. Proof (Sketch [8, Theorem IV.15]). This relies on the fact that satur (·) is computable in polynomial time and that the cyclicity of F and that of satur (F ) are related. More precisely F is cyclic iff either satur (F )↑ or satur (F )↓ is. Finally one has to see that the case of increasing or decreasing wirings is easy to treat by discarding the bottom of large stacks, which is harmless in that case. The saturation technique can then be used to show that the language recognized by an observation with stack always belongs to the class Ptime. The important point in the proof is that, given an observation O and a representa¯ p of a word W , one can produce in polynomial time an element of Stack tion W ¯ p. whose nilpotency is equivalent to the nilpotency of OW Proposition 1. Let O ∈ Ob+s be an observation with stack. There is a procedure RedO (·) ∈ FPtime that, given a word W as an input, outputs a wiring F ∈ Stack ¯ p is for any choice of p. with h(F ) ≤ h(O) such that F is nilpotent iff OW This is done essentially by remarking that the “balanced” part of the computation can never step outside a finite computation space, so that one can associate to each configuration a unary function symbol that is put on top of the stack. Theorem 6 (soundness). If O ∈ Ob+s , then L(O) ∈ Ptime.

5

Unary Logic Programming

In previous sections, we showed how the Stack semiring captures polynomial time computation. As we already mentioned, the elements of this semiring correspond to a specific class of logic programs, so that our results have a reading in terms of complexity of logic programming [22] which we detail now.

386

C. Aubert et al.

Definition 20 (Data, Goal, Query). A unary query is Q = (D, P, G), where: – D is a set of closed unary terms (a unary data), – P is a an element of Stack (a unary program), – G is a closed unary term (a unary goal). We say that the query Q succeeds if G  can be derived combining d  with d ∈ D and the elements of P by the resolution rule presented in Sect. 2.1, otherwise we say the query fails. The size |Q| of the query is defined as the total number of occurrences of symbols in it. To apply the saturation technique directly, we need to represent all the elements of the unary query (data, program, goal) as elements of Stack . This requires a simple encoding. Definition 21 (Encoding Unary Queries). We suppose that for any constant symbol c, we have a unary function symbol c(·). We also need two unary functions, START(·) and ACCEPT(·). To any unary data D we associate an element of Stack : [D] := {τ (c(x))  START(x)|τ (c) ∈ D} and to any unary goal G = τ (c) we associate G := ACCEPT(x)  τ (c(x)). Note that the program part P of the query needs not to be encoded as it is already an element of Stack . Once a query is encoded, we can tell if it is successful or not using the language of the resolution semiring. Lemma 2 (success). A unary query Q = (D, P, G) succeeds if and only if ACCEPT(x)  START(x) ∈ G P n [D] for some n. The saturation technique then can be applied to unary queries adding to new shortcut rules which eventually allow to decide acceptance. Lemma 3 (saturation of unary queries). A unary  query Q = (D, P, G) succeeds if and only if ACCEPT(x)  START(x) ∈ satur [D] + P + G . Theorem 7 (PTIME-completeness). The UQuery problem (given a unary query, is it successful?) is Ptime-complete. Proof. The lemma above, combined with the fact that satur (·) is computable in polynomial time3 , ensures that the problem lies indeed in the class Ptime. The hardness part follows from a variation on the encoding presented in Sect. 3 and the reduction derived from Proposition 1. Remark 3. We presented the result in a restricted form to stay in line with the previous sections. However, it should be clear to the reader that it would not be impacted if we allowed: non-closed goals and data; programs with no restriction on variables, e.g. f(x)  g(y); constants in the program part of the query. 3

The bound on the running time of the procedure computing satur (·) being exponential in the height, one needs to first process the query into an equivalent one using only terms of bounded height, which can easily be done in polynomial time.

Unary Resolution: Characterizing PTIME

387

Remark 4. In terms of complexity of logic programs, we are considering the combined complexity [22, p. 380]: every part of the query Q = (D, P, G) is variable. If for instance we fixed P and G (thus considering data complexity), we would have a problem that is still in Ptime, but it is unclear to us if it would be complete. Indeed, the encoding of Sect. 3 relies on a representation of inputs as plain programs, and on the fact that the evaluation process is a matter of interaction between programs rather than mere data processing.

6

Perspectives

Adding a “stack plugin” to observations extends modularly previous works [2, 3,5,6] and gives the perfect tool to characterize Ptime. This modularity was inspired by the classical addition of a stack to an automaton, allowing to switch from Logspace to Ptime, and providing a decisive proof technique: memoization—or exponentiation by squaring in our context—implemented as saturation. The automata’s qualitative constraint on memory is directly represented as a syntactic restriction on flows. In this setting, evaluation is inspired by the interactive approach to the Curry-Howard correspondence—geometry of interaction—, which makes the complexity parametric in the program and the input. This mechanism of computation differs from automata’s step-by-step evaluation, but that does not prevent the simulation of pushdown automata by unary logic program. The mechanism of pre-computation of transitions, known as memoization, was adapted in a setting where logic programs are represented as algebraic objects. This saturation technique computes shortcuts in a logic program to decide its nilpotency in polynomial time. As it turns out, this is similar to the techniques employed to solve efficiently the problem of termination of pushdown systems. More generally, this approach to complexity is based either on operator algebra [5, 6,31] or unification theory [2, 3, 8]: it is emerging as a meeting point for computer science, logic and mathematics, and raises a number of perspectives. A number of interrogations emerges naturally when considering the relations to proof theory. First, we could consider the Church encoding of other data types—trees for instance—and define “orthogonally” set of programs interacting with them, wondering what their computational nature is. In the distance, one may hope for a connection between our approach and ongoing work on higher order trees and model checking [33]; all alike, one could study the interaction between observations and one-way integers—briefly discussed in earlier work [3]—or non-deterministic data. Second, a still unanswered question of interest is to give proof-terms representation of captured programs, i.e. observations. Finally, it should be possible to represent functional computation (and not only decision problems, i.e. to switch from Ptime to FPtime), by considering a more general notion of observation that could express what an output is. In that perspective, a good place to start should be to show that light logics characterization results [9] can be recovered via our methods, which seems very likely but remains to be precisely investigated.

388

C. Aubert et al.

References 1. Aho, A.V., Hopcroft, J.E., Ullman, J.D.: Time and tape complexity of pushdown automaton languages. Inform. Control 13(3), 186–206 (1968) 2. Aubert, C., Bagnol, M.: Unification and logarithmic space. In: Dowek, G. (ed.) RTA-TLCA 2014. LNCS, vol. 8560, pp. 77–92. Springer, Heidelberg (2014) 3. Aubert, C., Bagnol, M., Pistone, P., Seiller, T.: Logic programming and logarithmic space. In: Garrigue, J. (ed.) APLAS 2014. LNCS, vol. 8858, pp. 39–57. Springer, Heidelberg (2014) 4. Aubert, C., Bagnol, M., Seiller, T.: Memoization for unary logic programming: Characterizing ptime. Research Report RR-8796, INRIA (2015). https://hal. archives-ouvertes.fr/hal-01107377 5. Aubert, C., Seiller, T.: Characterizing co-NL by a group action. In: MSCS (FirstView), pp. 1–33, December 2014 6. Aubert, C., Seiller, T.: Logarithmic space and permutations. Inf. Comput. Spec. Issue Implicit Comput. Complex. (2015). doi:10.1016/j.ic.2014.01.018 7. Baader, F., Nipkow, T.: Term rewriting and all that. CUP, Cambridge (1998) 8. Bagnol, M.: On the Resolution Semiring. Ph.D. thesis, Aix-Marseille Universit´e Institut de Math´ematiques de Marseille (2014). https://hal.archives-ouvertes.fr/ tel-01215334 9. Baillot, P., Mazza, D.: Linear logic by levels and bounded time complexity. Theoret. Comput. Sci. 411(2), 470–503 (2010) 10. Baillot, P., Pedicini, M.: Elementary complexity and geometry of interaction. Fund. Inform. 45(1–2), 1–31 (2001) 11. Baillot, P., Terui, K.: Light types for polynomial time computation in lambdacalculus. In: LICS, pp. 266–275. IEEE Computer Society (2004) 12. Bellantoni, S.J., Cook, S.A.: A new recursion-theoretic characterization of the polytime functions. Comput. Complex. 2, 97–110 (1992) 13. Carayol, A., Hague, M.: Saturation algorithms for model-checking pushdown sys´ tems. In: Esik, Z., F¨ ul¨ op, Z. (eds.), Proceedings 14th International Conference on Automata and Formal Languages, AFL 2014, Szeged, Hungary, May 27–29, 2014. EPTCS, vol. 151, pp. 1–24 (2014) 14. Caucal, D.: R´ecritures suffixes de mots. Research Report RR-0871, INRIA (1988). https://hal.inria.fr/inria-00075683 15. Caucal, D.: On the regular structure of prefix rewriting. In: Arnold, A. (ed.) CAAP ’90. LNCS, vol. 431, pp. 87–102. Springer, Heidelberg (1990) 16. Cook, S.A.: Characterizations of pushdown machines in terms of time-bounded computers. J. ACM 18(1), 4–18 (1971) 17. Cook, S.A.: Linear time simulation of deterministic two-way pushdown automata. In: IFIP Congress (1), pp. 75–80. North-Holland (1971) 18. Dal Lago, U.: A short introduction to implicit computational complexity. In: Bezhanishvili, N., Goranko, V. (eds.) ESSLLI 2010 and ESSLLI 2011. LNCS, vol. 7388, pp. 89–109. Springer, Heidelberg (2012) 19. Dal Lago, U., Sch¨ opp, U.: Functional programming in sublinear space. In: Gordon, A.D. (ed.) ESOP 2010. LNCS, vol. 6012, pp. 205–225. Springer, Heidelberg (2010) 20. Danos, V.: La Logique Lin´eaire appliqu´ee ´ a l’´etude de divers processus de normalisation (principalement du λ-calcul). Ph.D. thesis, Universit Paris VII (1990) 21. Danos, V., Joinet, J.B.: Linear logic and elementary time. Inf. Comput. 183(1), 123–137 (2003)

Unary Resolution: Characterizing PTIME

389

22. Dantsin, E., Eiter, T., Gottlob, G., Voronkov, A.: Complexity and expressive power of logic programming. ACM Comput. Surv. 33(3), 374–425 (2001) 23. Dwork, C., Kanellakis, P.C., Mitchell, J.C.: On the sequential nature of unification. J. Log. Program. 1(1), 35–50 (1984) 24. Gaboardi, M., Marion, J.Y., Rocca Della Rocca, S.: An implicit characterization of pspace. ACM Trans. Comput. Log. 13(2), 18:1–18:36 (2012) 25. Girard, J.Y.: Linear logic. Theoret. Comput. Sci. 50(1), 1–101 (1987) 26. Girard, J.Y.: Geometry of interaction 1: Interpretation of system F. Stud. Logic Found. Math. 127, 221–260 (1989) 27. Girard, J.Y.: Towards a geometry of interaction. In: Gray, J.W., Scedrov, A. (eds.) Proceedings of the AMS Conference on Categories, Logic and Computer Science. Categories in Computer Science and Logic, vol. 92, pp. 69–108. AMS (1989) 28. Girard, J.Y.: Geometry of interaction III: accommodating the additives. In: Girard, J.Y., Lafont, Y., Regnier, L. (eds.) Advances in Linear Logic, pp. 329–389. No. 222 in London Mathematical Society Lecture Note Series, CUP, June 1995 29. Girard, J.Y.: Light linear logic. In: Leivant, D. (ed.) Logic and Computational Complexity. LNCS, vol. 960, pp. 145–176. Springer, Heidelberg (1995) 30. Girard, J.Y.: Geometry of interaction V: Logic in the hyperfinite factor. Theoret. Comput. Sci. 412(20), 1860–1883 (2011) 31. Girard, J.Y.: Normativity in logic. In: Dybjer, P., Lindstr¨ om, S., Palmgren, E., Sundholm, G. (eds.) Epistemology versus Ontology. LEUS, vol. 27, pp. 243–263. Springer, Heidelberg (2012) 32. Gl¨ uck, R.: Simulation of two-way pushdown automata revisited. In: Banerjee, A., Danvy, O., Doh, K.G., Hatcliff, J. (eds.) Festschrift for Dave Schmidt. EPTCS, vol. 129, pp. 250–258 (2013) 33. Grellois, C., Melli´es, P.A.: Relational semantics of linear logic and higher-order model checking. In: Kreutzer, S. (ed.) CSL. LIPIcs, vol. 41, pp. 260–276. Schloss Dagstuhl - Leibniz-Zentrum fuer Informatik (2015). http://www.dagstuhl.de/ dagpub/978-3-939897-90-3 34. Huet, G., Lankford, D.: On the uniform halting problem for term rewriting systems. Research Report RR-283, INRIA (1978). http://www.ens-lyon.fr/LIP/ REWRITING/TERMINATION/Huet Lankford.pdf 35. Ladermann, M., Petersen, H.: Notes on looping deterministic two-way pushdown automata. Inf. Process. Lett. 49(3), 123–127 (1994) 36. Lafont, Y.: Soft linear logic and polynomial time. Theoret. Comput. Sci. 318(1), 163–180 (2004) 37. Leivant, D.: Stratified functional programs and computational complexity. In: Van Deusen, M.S., Lang, B. (eds.) POPL, pp. 325–333. ACM Press (1993) 38. Neergaard, P.M.: A functional language for logarithmic space. In: Chin, W.-N. (ed.) APLAS 2004. LNCS, vol. 3302, pp. 311–326. Springer, Heidelberg (2004) 39. Robinson, J.A.: A machine-oriented logic based on the resolution principle. J. ACM 12(1), 23–41 (1965) 40. Seiller, T.: Logique dans le facteur hyperfini : g´eometrie de l’interaction et complexit´e. Ph.D. thesis, Universit´e de la M´editerran´ee (2012), https://hal. archives-ouvertes.fr/tel-00768403 41. Seiller, T.: A correspondence between maximal abelian sub-algebras and linear logic fragments. ArXiv preprint abs/1408.2125, to appear in MSCS (2014) 42. Seiller, T.: Interaction graphs: Graphings. ArXiv preprint abs/1405.6331 (2014) 43. Wagner, K.W., Wechsung, G.: Computational Complexity, Mathematics and its Applications. Springer, Heidelberg (1986)

Focused and Synthetic Nested Sequents Kaustuv Chaudhuri(B) , Sonia Marin, and Lutz Straßburger ´ Inria & LIX/Ecole polytechnique, Palaiseau, France {kaustuv.chaudhuri,sonia.marin,lutz.strassburger}@inria.fr

Abstract. Focusing is a general technique for transforming a sequent proof system into one with a syntactic separation of non-deterministic choices without sacrificing completeness. This not only improves proof search, but also has the representational benefit of distilling sequent proofs into synthetic normal forms. We show how to apply the focusing technique to nested sequent calculi, a generalization of ordinary sequent calculi to tree-like instead of list-like structures. We thus improve the reach of focusing to the most commonly studied modal logics, the logics of the modal S5 cube. Among our key contributions is a focused cutelimination theorem for focused nested sequents.

1

Introduction

The focusing technique has its origin in the foundations of logic programming [1, 22] and is now increasingly relevant in structural proof theory because it improves proof search procedures [11,21] and because focused proofs have clearly identifiable and semantically meaningful synthetic normal forms [6, 8, 10, 31]. The essential idea of focusing is to identify and coalesce the non-deterministic choices in a proof, so that a proof can be seen as an alternation of negative phases, where invertible rules are applied eagerly, and positive phases, where applications of the other rules are confined and controlled. This, in turn, lets us abstract from the usual unary and binary logical connectives by collapsing whole phases into n-ary synthetic connectives. The full theory of focusing was initially developed for the sequent calculus for linear logic [1], but it has since been extended to a wide variety of logics [11, 19,27] and proof systems [4, 7]. This generality suggests that the ability to transform a proof system into a focused form is a good indication of its syntactic quality, in a manner similar to how admissibility of cut shows that a proof system is syntactically consistent. It is natural to ask whether the focusing technique works as well for modal logics. Traditionally, modal logics are specified in terms of Hilbert-style axiomatic systems, but such systems are not particularly suitable since axioms reveal none of the structure of logical reasoning. It is well known that certain modal logics, S5 in particular, are not representable in a variant of Gentzen’s sequent calculus without sacrificing analyticity. There are two principal ways to overcome this problem. The first is based on labeled proof systems that reify the Kripke semantics—the frame conditions—directly as formulas in the sequents [24, 29]. These “semantic formulas” are not subformulas of the end-sequent and they c Springer-Verlag Berlin Heidelberg 2016  B. Jacobs and C. L¨ oding (Eds.): FOSSACS 2016, LNCS 9634, pp. 390–407, 2016. DOI: 10.1007/978-3-662-49630-5 23

Focused and Synthetic Nested Sequents

391

cause the interpretation of sequents to fall outside the class of propositional modal formulas, and for this reason, such calculi are also called external. The second way is to use so-called internal calculi, that enrich the sequent structure such that analyticity is preserved and such that every sequent has an interpretation that stays inside the modal language. Well-known examples are hypersequents [2] and display calculi [3]. A more recent development are nested sequents [5,12,14,26], which generalize the notion of context from a listlike structure (familiar from Gentzen’s sequent calculus) to a tree-like structure. Like ordinary sequents, nested sequents have a straightforward interpretation in the language of the logic, and enjoy cut admissibility (with a cut-elimination proof that stays wholly internal to the system) and hence the usual subformula property. Moreover, nested proof systems can be built modularly for every modal logic in the S5 cube, in both classical and intuitionistic variants [5, 20]. In this paper, we build a focused variant, with its concomitant benefits, for all modal logics of the classical S5 cube. For simplicity we use a polarized syntax [15] consisting of two classes of positive and negative formulas and a pair of shift connectives to move back and forth between the classes. Crucially, we interpret 3 as positive and 2 as negative, which differs from the polarity that would be assigned to these connectives if they were interpreted in terms of ? and !, respectively, from linear logic [25]. Our key technical contributions are: (1) a purely internal proof of cut-elimination for the focused nested calculus, given in terms of a traditional rewriting procedure to eliminate cuts (which shows that our system is compositional and suitably continuous), and (2) a proof of completeness of the focused system with respect to the non-focused system (and hence to the Kripke semantics) by showing that the focused system admits the rules of the non-focused system. It generalizes similar proofs of cut-elimination and focusing completeness for (non-nested) sequent calculi [11, 19]. To our knowledge there have been only two other attempts to apply the focusing technique to modal logics. The first uses a labeled system [23], using the work in [24] on geometric axioms to obtain systems that extend the basic modal logic K. The cut-elimination and completeness results in [23] are obtained externally as a reduction to LKF, a focused system for first-order logic [19]. Therefore, the polarities of the modalities are inherited from the associated quantifiers, i.e., 2 is negative and 3 is positive, similar to our setting. The second approach [18] also uses nested sequents, but in a restricted form, in which the tree-structure is reduced to a single branch. This efficiently simulates the standard sequent system for the modal logic K, but makes both modalities positive. Our approach uses the full power of nested sequents. and is intended as a prototype for how similar focused systems may be built for other modal logic formalisms. After some preliminaries (Sect. 2), we start with a weakly focused proof system (Sect. 3), where negative rules may be applied everywhere, including in the middle of focused phases. From this system, we extract a strongly focused system (also Sect. 3) and a synthetic system (Sect. 4) where the logical content of the phases of focusing are abstracted from the level of formulas to the level of nested sequents. We also sketch the cut-elimination theorem for this synthetic variant. The synthetic design generalizes similar designs for the sequent calculus [6, 31].

392

K. Chaudhuri et al.

Fig. 1. Left: Some standard modal axioms

2

Right: Modal S5 cube

Modal Logics and the Nested Sequent Calculus KN

Classical modal logic is obtained from classical propositional logic by adding the modal connectives 2 and 3. Starting with a countable set of atoms (a, b, . . .), the formulas (A, B, . . .) of modal logic are given by the following grammar: A, B, ... ::= a |¯ a | A ∧ B | A ∨ B | 2A | 3A

(1)

To avoid excessive syntax, formulas are kept in negation-normal form, so the only formally negated formulas are the atoms. The negation A¯ of an arbitrary ¯ B ¯ and 2A = 3A. ¯ formula A is given by the De Morgan laws: A¯ = A, A ∧ B = A∨ ¯ We also define A ⊃ B as A ∨ B, A ≡ B as (A ⊃ B) ∧ (B ⊃ A),  as a ∨ a ¯, and ⊥ as a ∧ a ¯ (for some atom a). Modal logics are traditionally specified using Hilbert-style axiom schemata. The basic modal logic K, for instance, is obtained by adding the following k axiom to the ordinary Hilbert axioms for propositional logic. k : 2(A ⊃ B) ⊃ (2A ⊃ 2B)

(2)

To obtain the theorems of K, we then also add two inference rules of modus ponens and necessitation. mp

A A⊃B B

nec

A 2A

(3)

Stronger modal logics can be obtained by adding to K other axioms mentioning the modal connectives. In this paper, we consider the most common five axioms d, t, b, 4 and 5, which are shown on the left in Fig. 1. Picking subsets of these axioms lets us define thirty-two modal logics, but only fifteen of them are non-redundant. For example, the sets {b, 4} and {t, 5} both yield the modal logic S5. The fifteen distinct modal logics follow chains of extension from K to S5 and can be arranged as a pair of nested cubes depicted on the right in Fig. 1; this is sometimes called the S5 cube [13].

Focused and Synthetic Nested Sequents

393

Fig. 2. Rules for KN + X3 . The first row constitutes KN.

Let us recall the notion of nested sequents, first defined by Kashima [14] and then independently rediscovered by Poggiolesi [26] (who called them treehypersequents) and Br¨ unnler [5]. In Gentzen’s one-sided sequent calculus, a sequent is just a multiset of formulas; nested sequents generalize this notion to a multiset of formulas and boxed sequents, resulting in a tree structure. Definition 2.1. A nested sequent (Γ, Δ, . . .) is a finite multiset of formulas and boxed sequents of the form [Δ], where Δ is itself a nested sequent. In other words, nested sequents have the following grammar: Γ, Δ, . . . ::= A1 , . . . , Am , [Γ1 ], . . . , [Γn ]

(4)

Both m and n may be 0, in which case the sequent is empty; when we need to be explicit, we will use the notation ∅ to stand for an empty sequent. As is usual in sequent calculi, we consider the comma to be associative and commutative. Definition 2.2 (Corresponding Formulas). For any nested sequent Γ , a corresponding formula, written fm(Γ ), gives an interpretation of Γ as a modal logic formula. Corresponding formulas obey the following equivalences: fm(∅) ≡ ⊥, fm(A) ≡ A, fm([Γ ]) ≡ 2fm(Γ ), and fm(Γ1 , Γ2 ) ≡ fm(Γ1 ) ∨ fm(Γ2 ). Definition 2.3 (Context). An n-holed context is like a nested sequent but contains n pairwise distinct numbered holes of the form { }i (for 1 ≤ i ≤ n) in place of formulas. (No hole can occur inside a formula.) We depict such a context as Γ { }1 · · · { }n . Given such a context and n nested sequents Δ1 , . . . , Δn , we write Γ {Δ1 }1 · · · {Δn }n to stand for the nested sequent where the hole { }i (for 1 ≤ i ≤ n) in the context has been replaced by Δi , with the understanding that if Δi is empty then the hole is simply removed. Unless there is any ambiguity, we will omit the hole index subscripts in this paper to keep the notation light. Definition 2.4. The depth of Γ { }, written dp(Γ { }), is given inductively by: dp({ }) = 0, dp(Δ, Γ { }) = dp(Γ { }), and dp([Γ { }]) = dp(Γ { }) + 1. Example 2.5. Let Γ { }{ } = A, [B, { }, [{ }], C]. For the sequents Δ1 = D and Δ2 = A, [C], we get: Γ {Δ1 }{Δ2 } = A, [B, D, [A, [C]], C] and Γ {∅}{Δ2 } = A, [B, [A, [C]], C]. We also have that dp(Γ { }{Δ1 }) = 1 and dp(Γ {Δ1 }{ }) = 2.

394

K. Chaudhuri et al.

The basic modal logic K (as presented in [5]) is captured using nested sequents as the cut-free proof system KN shown in the first row in Fig. 2. The deductive system corresponding to each normal extension K + X, where X ⊆ {d, t, b, 4, 5} is a set of modal axioms (Fig. 1), can be obtained by adding the corresponding diamond rules X3 ⊆ {d3 , t3 , b3 , 43 , 53 } (final two rows of Fig. 2) to KN. The 53 rule has a side condition that the context in which the principal 3-formula occurs has non-zero depth, i.e., that it does not occur at the root of the conclusion sequent. To make the correspondence between extensions of K and the proof systems precise, we need the following additional notion: Definition 2.6 (45-Closure). We say that X ⊆ {d, t, b, 4, 5} is 45-closed, if: – whenever 4 is derivable in K + X, 4 ∈ X – whenever 5 is derivable in K + X, 5 ∈ X. In this paper, we will always work with 45-closed axiom sets. The reason is that, for example, the axiom 4 is not provable (without cut) in KN + {t3 , 53 } even though 4 is a theorem of the logic K + {t, 5} (which is S5). Note that this is not a real restriction, since for every logic in the modal cube (Fig. 1) there is a 45-closed set of axioms defining it (see [5] for details). We can now state the soundness, completeness, and cut-admissibility for KN and its extensions. Theorem 2.7. Let cut be the following rule: cut

¯ Γ {A} Γ {A} Γ {∅}

Let X ⊆ {d, t, b, 4, 5} be 45-closed. For any formula A, the following are equivalent. 1. A is a theorem of K + X. 2. A is provable in KN + X3 + cut. 3. A is provable in KN + X3 . The proof that 1 =⇒ 2 =⇒ 3 =⇒ 1 can be found in [5].

3



The Focused Systems KNwF and KNF

The essence of the focusing technique [1] is to classify formulas into positive formulas, whose rules are not invertible, and negative whose rules are invertible. (As usual, we consider a rule to be invertible if whenever the conclusion of the rule is derivable then so are each of its premises.) Due to invertibility, when searching for a proof it is always safe to apply—reading from conclusion to premises—a rule for a negative formula, so these may be applied at any time. On the other hand, rules for positive formulas may require rules on other formulas to be applied first. For example, the KN sequent 2¯ a, 3a can only be proved by first applying the 2 rule, showing that 3-formulas are positive. A focused proof is one where the decision to apply a rule to a positive formula has to be explicitly taken, which then commits the proof

Focused and Synthetic Nested Sequents

395

to continue applying rules to this focused positive formula and its immediate positive descendants (and no other formula in the sequent), which drastically reduces the search space. The main theorem of focusing is that this strategy is complete, i.e., every theorem has a focused proof. We will now build such a focused version of KN. To simplify the meta-theorems about this system, we will adopt a polarized syntax [15] where the positive and negative formulas are grouped together in different syntactic categories and explicitly mediated by shift connectives (↑ and ↓). As already mentioned, 3 is in the positive class, and its dual 2 is in the negative class, unlike what would be expected if they were interpreted in terms of the linear logic modalities ? and ! respectively. One way of explaining this phenomenon is that the standard shallow rule k in sequent calculus (corresponding to the promotion rule of linear logic) realises two steps at once: accessing the formula under a 2 and moving the 3-formulas from the context next to that formula. These two steps are done by two different rules 2 and k3 in nested sequent systems like KN. The rest of the formulas have ambiguous polarities and the choice does not alter the focusing result — some reasons to pick certain polarizations can be found in [11, 19]. We arbitrarily assign all atoms to be positive (and their negations to be negative), and present the system in the strongest form, i.e., split the conjunctions and disjunctions into positive and negative versions. Thus, polarized formulas have the following grammar: positive: negative:

P, Q, . . . ::= a | P ∧ Q | P ∨ Q | 3P | ↓N N, M, . . . ::= a ¯ | N ∨ M | N ∧ M | 2N | ↑P +



+

(5)



Each column in the grammar above defines a De Morgan dual pair; note that the negation of a positive formula is a negative formula, and vice versa. Units are a, f + = a ∧ ↓¯ a, definable similarly to the non-focused case, so we have t+ = a ∨ ↓¯ − − and dually for f /t . When the polarity of a formula is not important, we write it as A, B, . . .. A polarized nested sequent is the same as in the non-focused setting, with the difference that all formulas are polarized. Likewise, a polarized context is a polarized nested sequent where some formulas have been replaced by holes. In the rest of this paper, we will drop the adjective “polarized” and treat all constructs implicitly as polarized, unless otherwise indicated. +

+

Definition 3.1 (Neutral). A formula is said to be neutral if it is a positive formula or a negated atom. A nested sequent is neutral if it is built from multisets of neutral formulas and boxed neutral sequents. A context Γ { } · · · { } is neutral if Γ {∅} · · · {∅} is neutral. Definition 3.2. A focused sequent is of the form Γ {P } where Γ { } is a context and P is a positive formula. The formula P is called its focus. The notion of corresponding formula (Definition 2.2) is extended with fm(P ) ≡ fm(P ). The inference rules of the focused system KNwF (w for “weak”) are shown in the first three rows in Fig. 3 (the basic system for K). Then, for a set X ⊆ {d, t, b, 4, 5}, we write X3f ⊆ {k3f , t3f , b3f , 43f , 53f } for the corresponding subset

396

K. Chaudhuri et al.

Fig. 3. Focused Rules for KNF + X3f and KNwF + X3f . The first three rows constitute KNwF and KNF.

of the focused diamond rules in the last two rows. Observe that the rules for negative formulas are exactly the same as in the unfocused system, while the rules for positive formulas can only be applied if the principal formula is in focus. Mediating between ordinary and focused sequents are the rules dec (“decide”), that chooses a positive formula in the conclusion and focuses on a copy of it in the premise, and rel (“release”) that drops the focus on a shifted formula. Since dec keeps the original positive formula in the context, there is no need to incorporate contraction in every positive rule, like in KN. The sto (“store”) rule removes a shift in front of a positive formula and is used to produce neutral premises from non-neutral conclusions. We define the system KNF to be a restriction of KNwF where the conclusion of the dec rule is required to be neutral, as well as the contexts surrounding the focus in all rules involving focused sequents. Thus, in KNF, the dec rule is only applicable when no other rule is applicable (no negative rule as there is no negative formula in a neutral sequent and no positive rule as there is no focus), and hence we sometimes call it strongly focused. We immediately have the following proposition: Proposition 3.3. Let X ⊆ {d, t, b, 4, 5}. A formula A is provable in KNF + X3f if and only if it is provable in KNwF + X3f . Proof. A derivation in KNF + X3f is by definition also a derivation in KNwF + X3f . Conversely, to convert a derivation in KNwF + X3f into one in KNF + X3f , we first have to replace all instances of id with a sequence of applications of {2, ∧, ∧, sto} followed (reading from conclusion upwards) by id, to ensure that the conclusion of the id rule is neutral. Then, the negative rules {∧, ∨, 2} can be permuted down by straightforward rule permutations to ensure that dec only applies to neutral sequents. 







+

Focused and Synthetic Nested Sequents

397

Fig. 4. Structural rules.

Fig. 5. Structural modal rules for axioms d, t, b, 4, 5 (where dp(Γ { }{[Δ]}) ≥ 1 in 5[ ] )

In order to establish the soundness and completeness of KNwF, we use the obvious forgetful injection of the polarized syntax into the unpolarized syntax. Definition 3.4 (Depolarization). If A is a polarized formula, then we write A for the unpolarized formula obtained from A by erasing the shifts ↑ and ↓, collapsing ∧ and ∧ into ∧, and collapsing ∨ and ∨ into ∨. +



+



Theorem 3.5 (Soundness). Let X ⊆ {d, t, b, 4, 5}. If a formula A is provable in KNF + X3f , then A is provable in KN + X3 . Proof. By forgetting the polarity information, every KNF + X3f proof of A is

 transformed into a KN + X3 proof of A. Completeness is considerably trickier. We use a technique pioneered by Laurent for linear logic [16] and proceed via cut-elimination in KNwF. 3.1

Cut Elimination

In this section we will show that a collection of cuts is admissible for KNwF + X3f . As usual, a rule is said to be admissible if it is the case that whenever any instance of all its premises is derivable, so is the corresponding instance of the conclusion. In order to show the admissibility of the cut rules, it will be very useful to appeal to a collection of other admissible and invertible rules. Lemma 3.6. Let X ⊆ {d, t, b, 4, 5}. The rules weak, weakf , cont, and m[ ] (shown in Fig. 4) are admissible for KNwF + X3f . Moreover, the rules sto, 2, ∧, and ∨ are invertible for KNwF + X3f . −

Proof. By straightforward induction on the height of the derivation.





Note that here we use contraction only on negated atoms because that is all that is needed in the cut-elimination proof below. One can indeed show that the general contraction rule on arbitrary sequents (and not just formulas) is admissible, but this requires a complicated argument for focused sequents. The corresponding result for KN is shown in [5]. Lemma 3.7. Let X ⊆ {d, t, b, 4, 5}. If X is 45-closed, then any rule x[ ] in X[ ] (shown in Fig. 5) is admissible for KNwF + X3f .

398

K. Chaudhuri et al.

Fig. 6. The various cuts in KNwF + X3f

Proof. Analogous to the corresponding lemma for KN in [5, Lemma 9]. The full proof can be found in [9, Lemma 3.7]. 

We are now ready to prove the admissibility of cuts. Specifically, we show all the cuts in Fig. 6 are simultaneously admissible. The cut1 rule is our standard cut between ordinary nested sequents, while cut2 defines a principal cut between a focus and its dual as cut formulas. Finally, cut3 is a commutative cut for situations where the positive cut formula is not principal. Note that this collection of cuts is just sufficiently large to make the standard cuts admissible. It is easy to imagine many other cut-like rules, but it is not necessary—and may not even be possible—to admit them. Definition 3.8. The height of a formula A, written ht(A), is computed inductively as follows: ht(a) = ht(¯ a) = 1, ht(A  B) = max(ht(A), ht(B)) + 1 where  ∈ {∧, ∧, ∨, ∨}, and ht(A) = ht(A) + 1 where  ∈ {3, 2, ↑, ↓}. The rank of an instance of one of the cut rules is the height of its cut formula (the P in Fig. 6). +



+



Lemma 3.9 (Cut Reduction). Let X ⊆ {d, t, b, 4, 5} be 45-closed. For every derivation ?? ? ??D1  ???D2  ? ? (6) Γ1 Γ2 −−−−−−−−−−−−− − cuti − Γ0

in KNwF+X3f +{cut1 , cut2 , cut3 }, where D1 and D2 are cut-free, there is a cut-free derivation of Γ0 in KNwF + X3f . Proof. Let D1 always stand for the derivation with the positive cut formula. We proceed by lexicographic induction: the induction hypothesis may be applied whenever (1) the rank of the cut decreases, or (2) the rank stays the same and a cut1 is replaced by a cut2 , or (3) the rank stays the same and the height of D1 decreases. The height of D2 does not matter for the induction. The proof is then given in terms of a terminating rewrite sequence, written with ;. Most cases of this rewrite are standard, so we show here only certain cases characteristic of focusing; the full list of cases can be found in [9, Lemma 3.9].

Focused and Synthetic Nested Sequents

399

– The commutative cases are simple, relying on the invertibility of the negative rules (Lemma 3.7). Here we show the case of rel above a cut3 : D2 D1

D1

Γ {N }{P } rel −−−−−−−−−−−−−−−− Γ { ↓N  }{P } Γ {∅}{P¯ } cut3 −−−−−−−−−−−−−−−−−−−−−−−−−−−−−− Γ { ↓N  }{∅} D2

cut1

;

Γ {N }{P }

Γ {∅}{P¯ } weak −−−−−−−−−−¯−− Γ {N }{P }

− −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−− −

Γ {N }{∅} rel −−−−−−−−−−−−−−− Γ { ↓N }{∅}

The resulting cut2 can be reduced because it has a smaller height. – The cases of dec with the cut formula being principal: D1

D2

Γ {P, P }

Γ {P¯ }

D1

Γ {P, P  }

D2

dec −−−−−−−−−−−− Γ {P } Γ {P¯ } cut1 −−−−−−−−−−−−−−−−−−−− Γ {∅}

cut3

;

− −−−−−−−−−−−−−−−−−−−−− −

cut2

Γ { P }

D2

Γ {P¯ }

− −−−−−−−−−−−−−−−−−−−−−−−−−− −

Γ {∅}

In the resulting derivation, we first reduce the upper cut3 , which is allowed because the height is smaller. Then, we reduce the lower cut, which is allowed because a cut2 can be used to justify a cut1 of the same rank. – Finally, here is a characteristic case for the modal axioms, for the 43f rule: D2

D1

43 cut2

Γ {[3P  , Δ]}

D2

− −−−−−−−−−−−−−− −

Γ { 3P , [Δ]} Γ {2P¯ , [Δ]}

− −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−− −

Γ {[Δ]}

;

cut2

Γ {2P¯ , [Δ]} 2−1 ........................ Γ {[P¯ ], [Δ]} 4[ ] .......................  D1 Γ {[[P¯ ], Δ]} −−−−− − 2 −−−−−−−− Γ {[3P  , Δ]} Γ {[2P¯ , Δ]}

− −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−− −

Γ {[Δ]}

The resulting cut2 can be reduced because it has a smaller height. In the right branch, the 2−1 rule is the admissible inverse of the 2 rule (Lemma 3.6) and 

the 4[ ] rule is given by Lemma 3.7. Theorem 3.10. Let X ⊆ {d, t, b, 4, 5} be 45-closed. If a sequent Γ is provable in KNF + X3f + {cut1 , cut2 , cut3 }, then it is also provable in KNF + X3f . Proof. Apply Lemma 3.9 to all cut instances in the derivation, starting with a topmost one. This gives us a cut-free proof in KNwF + X3f , from which we get a 

cut-free proof in KNF + X3f using Proposition 3.3. 3.2

Completeness

We can now use Theorem 3.10 to show completeness of the focused systems KNF + X3f (and hence KNwF + X3f ) with respect to KN + X3 . As an intermediate step, we consider a variant of KN that can deal with polarized formulas. Let KN denote the system that is obtained from KN by adding the rules rel

Γ {N } Γ {↓N }

sto

Γ {P } Γ {↑P }

400

K. Chaudhuri et al.

and by duplicating the rules for ∧ and ∨ such that there is a variant for each of ∧ and ∧, and ∨ and ∨, respectively. We immediately have the following lemma: +



+



Lemma 3.11. A formula A is provable in KN +X3f if and only if A is provable 

in KN + X3 . We are now going to simulate KN + X3f in KNwF + X3f . For this, we need another property of KNwF: Lemma 3.12 (Identity Reduction). The following rule is derivable in KNwF: gid

Γ {P , P¯ }

Proof. By straightforward induction on the height of P . Details can be found in [9, Lemma 3.12]. 

Lemma 3.13 (Simulation). Let X ⊆ {d, t, b, 4, 5} and let A be a formula that is provable in KN + X3f . Then A is provable in KNwF + X3f + cut1 . Proof. We show that each rule in KN + X3f is derivable in KNwF + X3f + cut1 . Then the lemma follows by replacing in the proof of A in KN + X3f each instance of a rule by the corresponding derivation in KNwF + X3f . For the rules sto, 2, ∧, ∨ this is trivial. We show below the cases for ∨, d3 , and 43 . The others are similar, and a full list can be found in [9, Lemma 3.13]. −



+

gid ======================¯== Γ {P ∨ Q, Q, P, Q} −−−−−−−−−−−−−−−−−−−−−−−−−− − gid ===================¯== ∨− ¯ Γ {P ∨ Q, P , P } Γ {P, Q} Γ {P ∨ Q, P ∨ Q, P, Q} ............................... −−−−−−−−−−−−−−−−−−−−−−− − − − − − − − − − − − − − − − − − − − − − − − − − − − − − ∨− weak dec ¯ Γ {P ∨ Q, P, Q} Γ {P ∨ Q, P, Q} Γ {P ∨ Q, P ∨ Q, P¯ } − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − cut1 dec Γ {P ∨ Q, P } Γ {P ∨ Q, P¯ } − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − cut1 Γ {P ∨ Q} +

+

+

+

+

+

+

+

+

+

+

+

+

gid =================¯========¯== Γ {3P, [P , P ], 3↓P } − −−−−−−−−−−−−−−−−−−−−−−−−−− − Γ {3P, 3P , [P¯ ], 3↓P¯ } −−−−−−−−−−−− − dec −−−−−−−−−−−−−−− Γ {3P, [P¯ ], 3↓P¯ } rel −−−−−−−−−−−−−−−−−−−−−−−− Γ {3P, [↓P¯ ], 3↓P¯ } d3f −−−−−−−−−−−−−−−−−−−−−−−−− Γ {3P, 3↓P¯ , 3↓P¯ } dec −−−−−−−−−−−−−−−−−−¯−−−−−−− Γ {3P, 3↓P } k3f

Γ {3P, [P ]} sto −−−−−−−−−−−−−−− Γ {3P, [↑P ]} 2 −−−−−−−−−−−−−−− Γ {3P, 2↑P } cut1 −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−− Γ {3P }

===== = gid ====================== Γ {3P, [3P , 2P¯ , Δ]} −−−−− − 43f −−−−−−−−−−−−−−−−−−−−− Γ {3P, 3P , [2P¯ , Δ]} −−−−−−−−− − dec −−−−−−−−−−−−−−−−− Γ {3P, [2P¯ , Δ]}

Γ {3P, [3P, Δ]} cut1 −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−− Γ {3P, [Δ]}

Note that we make a crucial use of cut1 and Lemma 3.12.



Theorem 3.14 (Completeness). Let X ⊆ {d, t, b, 4, 5} be 45-closed. For any A, if A is provable in KN + X3 , then A is provable in KNF + X3f . Proof. Suppose that we have a proof of A in KN + X3 . By Lemma 3.11, there is a proof of A in KN + X3f . Then, by Lemma 3.13, there is also a proof of A in  KNwF + X3f + cut1 . Finally, using Theorem 3.10, we get a proof in KNF + X3f .

Focused and Synthetic Nested Sequents

4

401

The Synthetic System

As already mentioned, the strongly focused system KNF is given as a restriction of KNwF where the dec rule is restricted to neutral contexts. However, the cutelimination and admissibility theorems (3.9 and 3.10) were proved in the KNwF system and made essential use of the admissibility of weakening by arbitrary formulas, including negative formulas, and of the possibility of applying negative rules even within a positive phase. This freedom simplifies the proofs of the metatheorems, and leaves them at least a recognizable variant of similar proofs in the non-focused system KN. Of course, thanks to Proposition 3.3, we also have a cutelimination proof for KNF + X3f , but this is not entirely satisfactory: it is not an internal proof, i.e., a sequence of cut reductions for KNF + X3f + {cut1 , cut2 , cut3 } that stays inside the system. One possible response to this issue might be to try to redo the cut-elimination using just KNF + X3f , but this quickly gets rather complicated because we no longer have access to the weakening rules (Fig. 4) in the case where the weakened structure contains negative formulas. Indeed, published proofs of similar attempts for the sequent calculus usually solve this problem by adding additional cut rules, which greatly complicates the cut-elimination argument [11, 19, 28].

Fig. 7. Synthetic substructure matching

To avoid this complexity, it is better to consider the focused proof system in a synthetic form where the logical inference rules for the various connectives are composed as much as possible, so that the proof system itself contains exactly two logical rules: one for a positive and one for a negative synthetic connective [6,31]. This synthetic view moreover improves the concept of focusing itself by showing that a focused proof consists of: (1) a selection of a certain substructure (the generalization of subformula) of the principal formula, and (2) the contextualization of that substructure. For positive principal formulas, this contextualization is in the form of a check in the surrounding context for other structures, such as dual atoms or nested sequents. For negative principal formulas, on the other hand, contextualization amounts to asserting the presence of additional structure in the surrounding context. This design will become clear in the explicit formulation of the synthetic version of KNF—which we call KNS—in the rest of this section.

402

4.1

K. Chaudhuri et al.

Synthetic Substructures

For any negative formula, there is a collection of corresponding nested sequents that represents one of the possible branches taken in a sequence of negative rules applied to the formula. This correspondence is formally given below. Definition 4.1 (Matching). The nested sequence Γ matches the negative formula N , written Γ  N , if it is derivable from the rules in Fig. 7. For the system to follow, we will use two additional sequent-like structures that are not themselves neutral sequents. Definition 4.2 (Extended Sequents) – An inversion sequent is a structure of the form Γ {N } where Γ { } is a neutral sequent context. – A focused sequent is a structure of the form Γ {Δ} where Γ { } is a neutral sequent context and Δ is a neutral sequent. Note that in Γ {N }, there is exactly one occurrence of N as a top-level formula anywhere; likewise, in Γ {Δ}, there is a single occurrence of the sub-structure Δ. Hence, these extended sequent forms uniquely determine their decomposition into context (the Γ { }) and the extended entity (the N or the Δ). Definition 4.3 (Corresponding Formula). A corresponding formula of a neutral or extended sequent Γ , denoted as f m(Γ ¯ ), is a negative formula satisfying: f m(P ¯ ) ≡ ↑P f m(N ¯ )≡N   f m(Δ) ¯ ≡ ↑ f m(Δ) ¯

¯ f m(Γ, ¯ Δ) ≡ f m(Γ ¯ ) ∨ f m(Δ) −

f m([Γ ¯ ]) ≡ 2f m(Γ ¯ ) ¯ where N ≡ M stands for (↑N





¯ M ) ∧ (↑M −





N ).

The system KNS consists of the rules in the first two lines of Fig. 8. For a set X ⊆ {d, t, b, 4, 5}, we write X ⊆ {d , t , b , 4 , 5 } for the corresponding structural rules in Fig. 8, and KNS+X for the corresponding system. It becomes clearer that the duality between positive and negative synthetic rules amounts to a meta-quantification over substructures: the positive rule pos quantifies existentially over the substructures of P¯ and pick one such Δ as a focus in the unique premise, while the negative rule quantifies universally, and so the rule actually consists of one premise for each way in which to prove Δ  N . Thus, ¯  N and [¯b, P ]  N , so for example, if N is a ¯ ∧ 2(¯b ∨ ↑P ), then we know that a the rule instance in this case is: −



neg

Γ {¯ a} Γ {[¯b, P ]} Γ {¯ a ∧ 2(¯b ∨ ↑P )} −



It is instructive to compare KNS + X with KNwF + X3f (and hence also KNF + X3f ). In the latter system, the focus 3P  is used to drive the modal rules

Focused and Synthetic Nested Sequents

403

Fig. 8. Synthetic rules for KNS + X . The first two rows constitute KNS.

{d3f , k3f , t3f , b3f , 43f , 53f }. Such modal rules can be applied only a fixed number of times before 3P  needs to be reduced to P , and logical rules or the identity to be used — which is necessary to finish the proof since foci can never be weakened. Thus, the analysis of P is forced to be interleaved with the modal rules for 3P , as shown by the alternation of d3f and ∨ in the derivation on the left below. In KNS + X , in contrast, the pos rule itself performs the analysis of P up front to produce a synthetic substructure; the modal rules {k , d , t , b , 4 , 5 } then work entirely at the level of focused substructures, as in the derivation on the right below. +

id ======================= 3(p ∨ 3p), [[p, p¯]]

id

+

3

− −−−−− −−−−−−−−−−−−−−−−− − + 3(p ∨ 3p), [3p, [¯ p]] + −−−−− −−−−−−−−−−−− −−−−−−−−− − ∨− + + 3(p ∨ 3p), [p ∨ 3p, [¯ p]] 3 − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − −− − kf + + 3(p ∨ 3p), 3(p ∨ 3p), [[¯ p]] dec −−−−−−−−−−−−−−+−−−−−−−−−−−−−−−−−−− 3(p ∨ 3p), [[¯ p]]

kf

= ===================== = + 3(p ∨ 3p), [[¯ p, p¯]] − −−−−−+−−−−−−−−−−−−−−−−− − 3(p ∨ 3p), [[¯ p] , [¯ p]]  k −−−−−−+−−−−−−−−−−−−−−−−−−− + [[¯ p]]  3(p ∨ 3p) 3(p ∨ 3p), [[¯ p]] , [[¯ p]] − −−−−−−−−−−−−−−−−−−−−−−− −−−−−−−−−−−−−−−−−−−−−−−−−−− − + 3(p ∨ 3p), [[¯ p]]

k

pos

Thus, the modal rules of KNS + X are properly seen as structural rules rather than logical rules. 4.2

Cut Elimination

Cut elimination for KNS+X is achieved in a similar way to that for KNwF+X3f . Lemma 4.4 (Admissible Rules). The rules weak, weakf , cont, m[ ] (shown in Fig. 4), restricted to neutral and extended sequents (as appropriate) are admissible in KNS + X . Moreover, if X ⊆ {d, t, b, 4, 5} is 45-closed, then any rule x[ ] ∈ X[ ] (see Fig. 5) is admissible in KNS + X . Proof. By induction on the height of the derivation, analogous to the proofs of Lemmas 3.6 and 3.7.

 Note that we do not allow, e.g., weakening Γ {Δ}{∅} to Γ {Δ}{N }; the latter is, in fact, not even a well-formed KNS focused sequent. Like with KNwF,

404

K. Chaudhuri et al.

Fig. 9. Synthetic variants of the cut rule (cf. Fig. 6)

we have three cut rules for KNS, which are shown in Fig. 9. We can now give the synthetic variant of the cut-elimination theorem. Theorem 4.5. Let X ⊆ {d, t, b, 4, 5} be    {cut1 , cut2 , cut3 } is admissible in KNS + X .

45-closed.

Each

rule

of

Proof. The idea of the proof is similar to that of Theorem 3.10, using a reduction similar to that of Lemma 3.9. However, in the synthetic case there is a complication because the inverse of the negative rule neg is not admissible in KNS + X . The cut reduction argument therefore has to work at the level of synthetic derivations. We show two illustrative examples here. The first is the case where the cut formula is principal in both derivations: ⎧ ⎪ ⎨ D

pos

Δ  P¯

Γ {P, Δ }

− −−−−−−−−−−−−−−−−−−−−− −

cut 1

Γ {P }

⎪ ⎩

neg

⎫ ⎪ ⎬ DΔ

⎪ ⎭

Γ {Δ} ΔP¯ − −−−−−−−−−−−−−−−−− − Γ {P¯ }

cut 3 ;

− −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−− −

Γ {∅}

D

D2

Γ {P, Δ }

Γ {P¯ }

− −−−−−−−−−−−−−−−−−−−−− −

cut 2

Γ { Δ }



Γ {Δ}

− −−−−−−−−−−−−−−−−−−−−−−−−−−− −

Γ {∅}

 cut3

The instance of can be reduced because the height is smaller, while    that of cut2 can be reduced because a cut2 can justify a cut1 of the  same rank. The other illustrative case is for permuting a cut3 past a rel . ⎧ ⎪ ⎨

⎫ ⎪ ⎬ DΔ

⎪ ⎪ ⎩ ⎭ Γ {Δ}{P } ΔQ¯  − − − − − − − − − − − − − − − −−−−−− − neg ¯ D2 Γ {Q}{P }  rel −−−−−−−−−−−−−− neg Γ { Q}{P } Γ {∅}{ P¯ }  cut3 −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−− ; Γ { Q }{∅}

⎧ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎨

D2

⎫ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎬

DΔ Γ {∅}{P¯ } ⎪ ⎪ weak ..................... ⎪ ⎪ ⎪ ⎪ ⎪ Γ {Δ}{P } Γ {Δ}{P¯ }⎪ ⎪ ⎪  ⎪ ⎭ ⎩cut1 −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−⎪ Γ {Δ}{∅} ¯ ΔQ − −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−− − ¯ Γ { Q}{∅} rel −−−−−−−−−−−−− Γ {Q }{∅}

The instance of weak is justified by Lemma 4.4. The remainder of the cases of the proof have to be adjusted similarly. The full proof can be found in [9, Theorem 4.5]. 

It is worth remarking that this cut-elimination proof did not have to mention any logical connectives, and was instead able to factorize all logical reasoning in terms of the matching. This means that the matching judgement can be modified at will without affecting the nature of the cut argument, as long as it leaves the structure of nested sequents untouched. This makes our result modular in yet

Focused and Synthetic Nested Sequents

405

another way, in addition to the modularity obtained by means of the structural rules for foci. Indeed, we can obtain a similarly synthetic version of identity reduction (Lemma 3.12). Lemma 4.6 (Identity Reduction). The following rule is derivable in KNS. sid

Γ {Δ, Δ}

Proof. By induction on the structure of the focus, Δ. The full proof can be found in [9, Lemma 4.6].

 Showing KNS + X sound and complete with respect to KN + X3 is fairly straightforward and the details are omitted here. Soundness follows directly from replacing the KNS sequent Γ {Δ} with the KNF sequent Γ { f m(Δ) ¯ } and then interpreting the KNS + X proof in KNF + X3f , using matching derivation (Fig. 7) to determine how to choose between the two ∨ rules. For completeness, we can follow the strategy of Lemma 3.13 nearly unchanged. However, like with the cutelimination proof, we have to avoid appeals to weakening with negative formulas by using the synthetic form of neg . +

5

Perspectives

We have presented strongly focused and synthetic systems for all modal logics in the classical S5-cube. We used the formalism of nested sequents as carrier, but we are confident that something similar can be achieved for hypersequents, for example based on the work of Lellmann [17]. We chose nested sequents over hypersequents for two reasons. First, the formula interpretation of a nested sequent is the same for all logics in the S5-cube, which simplifies the presentation of the meta-theory. Second, due to the close correspondence between nested sequents and prefixed tableaux [12] we can from our work directly extract focused tableau systems for modal logics. Furthermore, even though we spoke in this paper only about classical modal logic, we are confident that the same results can also be obtained for the intuitionistic variant of the modal S5-cube [29], if we start from the non-focused systems in [30]. One extension that would be worth considering would be relaxing the restriction that there can be at most one focus in a KNF or KNS proof. Allowing multiple foci would take us from ordinary focusing to multi-focusing, which is well known to reveal more parallelism in sequent proofs [10]. It has been shown that a certain well chosen multi-focusing system can yield syntactically canonical representatives of equivalence classes of sequent proofs for classical predicate logic [8]. Extending this approach to the modal case seems promising.

406

K. Chaudhuri et al.

References 1. Andreoli, J.-M.: Logic programming with focusing proofs in linear logic. J. Logic Comput. 2(3), 297–347 (1992) 2. Avron, A.: The method of hypersequents in the proof theory of propositional nonclassical logics. In: Logic: From Foundations to Applications: European Logic Colloquium, pp. 1–32. Clarendon Press (1996) 3. Belnap Jr., N.D.: Display logic. J. Philos. Logic 11, 375–417 (1982) 4. Brock-Nannestad, T., Sch¨ urmann, C.: Focused natural deduction. In: Ferm¨ uller, C.G., Voronkov, A. (eds.) LPAR-17. LNCS, vol. 6397, pp. 157–171. Springer, Heidelberg (2010) 5. Br¨ unnler, K.: Deep sequent systems for modal logic. Arch. Math. Logic 48(6), 551–577 (2009) 6. Chaudhuri, K.: Focusing strategies in the sequent calculus of synthetic connectives. In: Cervesato, I., Veith, H., Voronkov, A. (eds.) LPAR 2008. LNCS (LNAI), vol. 5330, pp. 467–481. Springer, Heidelberg (2008) 7. Chaudhuri, K., Guenot, N., Straßburger, L.: The focused calculus of structures. In: Computer Science Logic: 20th Annual Conference of the EACSL. Leibniz International Proceedings in Informatics (LIPIcs), pp. 159–173. Schloss DagstuhlLeibniz-Zentrum f¨ ur Informatik, Sept 2011 8. Chaudhuri, K., Hetzl, S., Miller, D.: A multi-focused proof system isomorphic to expansion proofs. J. Logic Comput., June 2014. doi:10.1093/logcom/exu030 9. Chaudhuri, K., Marin, S., Straßburger, L.: Focused and Synthetic Nested Sequents. Technical report, Inria (2016). https://hal.inria.fr/hal-01251722 10. Chaudhuri, K., Miller, D., Saurin, A.: Canonical sequent proofs via multi-focusing. In: Ausiello, G., Karhum¨ aki, J., Mauri, G., Ong, L. (eds.) IFIP-TCS 2008. IFIP, vol. 273, pp. 383–396. Springer, Heidelberg (2008) 11. Chaudhuri, K., Pfenning, F., Price, G.: A logical characterization of forward and backward chaining in the inverse method. J. Autom. Reasoning 40(2–3), 133–177 (2008) 12. Fitting, M.: Prefixed tableaus and nested sequents. Ann. Pure Appl. Logic 163, 291–313 (2012) 13. Garson, J.: Modal logic. Stanford University, In The Stanford Encyclopedia of Philosophy (2008) 14. Kashima, R.: Cut-free sequent calculi for some tense logics. Stud. Logica. 53(1), 119–136 (1994) ´ 15. Laurent, O.: Etude de la Polarisation en Logique. Ph.D. thesis, Univiversit AixMarseille II (2002) 16. Laurent, O.: A proof of the focalization property in linear logic (2004) (Unpublished note) 17. Lellmann, B.: Axioms vs hypersequent rules with context restrictions: theory and applications. In: Demri, S., Kapur, D., Weidenbach, C. (eds.) IJCAR 2014. LNCS, vol. 8562, pp. 307–321. Springer, Heidelberg (2014) 18. Lellmann, B., Pimentel, E.: Proof search in nested sequent calculi. In: Davis, M., Fehnker, A., McIver, A., Voronkov, A. (eds.) LPAR-20 2015. LNCS, vol. 9450, pp. 558–574. Springer, Heidelberg (2015). doi:10.1007/978-3-662-48899-7 39 19. Liang, C., Miller, D.: Focusing and polarization in linear, intuitionistic, and classical logics. Theoret. Comput. Sci. 410(46), 4747–4768 (2009) 20. Marin, S., Straßburger, L.: Label-free modular systems for classical and intuitionistic modal logics. In: Advances in Modal Logic (AIML-10) (2014)

Focused and Synthetic Nested Sequents

407

21. McLaughlin, S., Pfenning, F.: Imogen: focusing the polarized inverse method for intuitionistic propositional logic. In: Cervesato, I., Veith, H., Voronkov, A. (eds.) LPAR 2008. LNCS (LNAI), vol. 5330, pp. 174–181. Springer, Heidelberg (2008) 22. Miller, D., Nadathur, G., Pfenning, F., Scedrov, A.: Uniform proofs as a foundation for logic programming. Ann. Pure Appl. Logic 51, 125–157 (1991) 23. Miller, D., Volpe, M.: Focused labeled proof systems for modal logic. In: Davis, M., Fehnker, A., McIver, A., Voronkov, A. (eds.) LPAR-20 2015. LNCS, vol. 9450, pp. 266–280. Springer, Heidelberg (2015). doi:10.1007/978-3-662-48899-7 19 24. Negri, S.: Proof analysis in modal logic. J. Philos. Logic 34(5–6), 507–544 (2005) 25. Pfenning, F., Davies, R.: A judgmental reconstruction of modal logic. Math. Struct. Comput. Sci. 11, 511–540 (2001). Notes to an invited talk at the Workshop on Intuitionistic Modal Logics and Applications (IMLA 1999) 26. Poggiolesi, F.: The method of tree-hypersequents for modal propositional logic. In: Makinson, D., Malinowski, J., Wansing, H. (eds.) Towards Mathematical Philosophy. Trends in Logic, vol. 28, pp. 31–51. Springer, New York (2009) 27. Reed, J., Pfenning, F.: Focus-preserving embeddings of substructural logics in intuitionistic logic. Draft manuscript, January 2010 28. Simmons, R.J.: Structural focalization. ACM Trans. Comput. Log. 15(3), 21:1–21:33 (2014) 29. Simpson, A.: The proof theory and semantics of intuitionistic modal logic. Ph.D thesis, University of Edinburgh (1994) 30. Straßburger, L.: Cut elimination in nested sequents for intuitionistic modal logics. In: Pfenning, F. (ed.) FOSSACS 2013 (ETAPS 2013). LNCS, vol. 7794, pp. 209–224. Springer, Heidelberg (2013) 31. Zeilberger, N.: Focusing and higher-order abstract syntax. In: Proceedings of the 35th ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages, POPL, San Francisco, California, USA, January 7–12, pp. 359–369. ACM (2008)

Strong Normalizability as a Finiteness Structure via the Taylor Expansion of λ-terms Michele Pagani1(B) , Christine Tasson1(B) , and Lionel Vaux2(B) 1

Universit´e Paris Diderot, CNRS, IRIF UMR 8243, 75205 Paris, France {michele.pagani,christine.tasson}@pps.univ-paris-diderot.fr 2 Aix Marseille Universit´e, CNRS, Centrale Marseille, I2M UMR 7373, 13453 Marseille, France [email protected]

Abstract. In the folklore of linear logic, a common intuition is that the structure of finiteness spaces, introduced by Ehrhard, semantically reflects the strong normalization property of cut-elimination. We make this intuition formal in the context of the non-deterministic λ-calculus by introducing a finiteness structure on resource terms, which is such that a λ-term is strongly normalizing iff the support of its Taylor expansion is finitary. An application of our result is the existence of a normal form for the Taylor expansion of any strongly normalizable non-deterministic λ-term.

1

Introduction

It is well-known that sets and relations can be presented as a category of modules and linear functions over the boolean semi-ring, giving one of the simplest semantics of linear logic. In [10] (see also [9]), it is shown how to generalize this construction to any complete1 semi-ring R and yet obtain a model of linear logic. In particular, the composition of two matrices φ ∈ RA×B and ψ ∈ RB×C is given by the usual matrix multiplication:  (φ; ψ)a,c : = φa,b · ψb,c (1) b∈B

The semi-ring R must be complete because the above sum might be infinite. This is an issue, because it prevents us from considering standard vector spaces, which are usually constructed over “non-complete” fields like reals or complexes. In order to overcome this problem, Ehrhard introduced the notion of finiteness space [3]. A finiteness space is a pair of a set A and a set A (called finiteness structure in Definition 7) of subsets of A which is closed under a notion of duality. The point is that, for any field K (resp. any semi-ring), the set of vectors in KA whose support2 is in A constitutes a vector space (resp. a module) over K. 1 2

Supported by French ANR Project Coquas (number ANR 12 JS02 006 01). A semi-ring is complete if any sum, even infinite, is well-defined. The support of v ∈ KA is the set of those a ∈ A such that the scalar va is non-null.

c Springer-Verlag Berlin Heidelberg 2016  B. Jacobs and C. L¨ oding (Eds.): FOSSACS 2016, LNCS 9634, pp. 408–423, 2016. DOI: 10.1007/978-3-662-49630-5 24

Strong Normalizability as a Finiteness Structure

409

Moreover, any two matrices φ ∈ KA×B and ψ ∈ KB×C whose supports are in resp. A  B and B  C (the finiteness structures associated with the linear arrow) compose, because the duality condition on the supports makes the terms in the sum of Eq. (1) be zero almost everywhere. This gives rise to a category which is a model of linear logic and its differential extension. The notion of finiteness space seems strictly related to the property of normalization. Already in [3], it is remarked that the coKleisli category of the exponential comonad is a model of simply typed λ-calculus, but it is not cpo-enriched and thus cannot interpret (at least in a standard way) fixed-point combinators, so neither PCF nor untyped λ-calculus. Moreover, in the setting of differential nets, Pagani showed that the property of having a finitary interpretation corresponds to an acyclicity criterion (called visible acyclicity [12]) which entails the normalization property of the cut-elimination procedure [11], while there are examples of visibly cyclic differential nets which do not normalize. The goal of this paper is to shed further light on the link between finiteness spaces and normalization, this time considering the non-deterministic untyped λ-calculus. Since we deal with λ-terms and not with linear logic proofs (or differential nets), we will speak about formal power series rather than matrices at the semantical level. This corresponds to move from the morphisms of a linear category to those of its coKleisli construction. Moreover, following [7], we describe the monomials of these power series as resource terms in normal form. The benefit of this setting is that the interpretation of a λ-term M as a power series [|M |] can be decomposed in two distinct steps: first, the term M is associated with a formal series M ∗ of resource terms possibly with redexes, called the Taylor expansion of M (see Table 1a); second, one reduces each resource term t appearing in the support T (M ) of M ∗ into a normal form NF(t) and sum up all the results, that is (Mt∗ denotes the coefficient of t in M ∗ ):  Mt∗ · NF(t) (2) [|M |] = t∈T (M )

The issue about the convergence of infinite sums appears in Eq. (2) because there might be an infinite number of resource terms in T (M ) reducing to the same normal form and thus possibly giving infinite coefficients to the formal series [|M |]. Ehrhard and Regnier proved in [7] that this is not the case for deterministic λ-terms, however the situation gets worse in presence of nondeterministic primitives. If we allow sums M +N representing potential reduction to M or N , then one can construct terms evaluating to a variable y an infinite number of times, like (where Θ denotes Turing’s fixed-point combinator): (Θ) λx.(x + y) →∗β (Θ) λx.(x + y) + y →∗β (Θ) λx.(x + y) + y + y →∗β . . . (3) We postpone to Examples 1 and 4 a more detailed discussion of (Θ) λx.(x + y), however we can already guess that this interplay between infinite reductions and non-determinism may produce infinite coefficients. One can then wonder whether there are interesting classes of terms where the coefficients of the associated power series can be kept finite. Ehrhard proved

410

M. Pagani et al.

Fig. 1. Main results of the paper

in [4] that the terms typable by a non-deterministic variant of Girard’s System F have always finite coefficients. A by-product of our results is Corollary 39, which is a generalization of Ehrhard’s result: every strongly normalizable nondeterministic λ-term can be interpreted by a power series with finite coefficients. The main focus of our paper is however in the means used for obtaining this result rather than on the result itself. The proof in [4] is based on a finiteness structure S over the set of resource terms Δ, such that any term M typable in System F has the support T (M ) of its Taylor expansion in S. We show that this method can be both generalized and strengthened in order to characterize strong normalization via finiteness structures. Namely, we give sufficient conditions on a finiteness structure S over Δ such that for every non-deterministic λ-term M : (i) if M is strongly normalizable, then T (M ) ∈ S (Corollary 30); (ii) if T (M ) ∈ S, then M is strongly normalizable (Theorem 36). Contents. Section 2 gives the preliminary definitions: the non-deterministic λ-calculus, its Taylor expansion into formal series of resource terms and the notion of finiteness structure. The proof of Item (i) splits into Sects. 3 and 4, using an intersection type system (Table 2) for characterizing strong normalization. Section 5 gives the proof of Item (ii) and Sect. 6 concludes with Corollary 39 about the finiteness of the coefficients of the power series of strongly normalizable terms. Figure 1 sums up the main results of the paper.

2

Preliminaries

2.1

Non-deterministic λ-calculus Λ+

The non-deterministic λ-calculus is defined by the following grammar3 : λ-terms Λ+ :

M : = x | λx.M | (M ) M | M + M

subject to α-equivalence and to the following identities: M + N = N + M,

(M + N ) + P = M + (N + P ),

λx.(M + N ) = λx.M + λx.N , (M + N ) P = (M ) P + (N ) P. 3

We use Krivine’s notation with the standard conventions, see [8] for reference.

Strong Normalizability as a Finiteness Structure

411

The last two equalities state that abstraction is a linear operation (i.e. commutes with sums) while application is linear only in the function but not in the argument (i.e. (P ) (M + N ) = (P ) M + (P ) N ). Notice also that the sum is not idempotent: M + M = M . This is a crucial feature for making a difference between terms reducing to a value once, twice, more times or an infinite number of times (see the discussion about Eq. (3) in the Introduction). Although this follows an old intuition from linear logic, the first extension of the λ-calculus with (a priori non idempotent) sums subject to these identities was, as far as we know, the differential λ-calculus of Ehrhard and Regnier [5]. This feature is now quite standard in the literature following this revival of quantitative semantics. The (capture avoiding) substitution of a term for a variable is defined as usual and β-reduction →β is defined as the context closure of: (λx.M ) N →β M [N/x] . We denote as →∗β the reflexive-transitive closure of →β . Example 1. Some λ-terms which will be used in the following are: Δ : = λx. (x) x,

Ω : = (Δ) Δ,

Δ3 : = λx. (x) xx,

Ω3 : = (Δ3 ) Δ3 ,

Θ : = (λxy. (y) (x) xy) λxy. (y) (x) xy. The term Ω is the prototypical diverging term, reducing to itself in one single β-step. Ω3 is another example of diverging term, producing terms of greater and greater size: Ω3 →β (Ω3 ) Δ3 →β (Ω3 ) Δ3 Δ3 →β . . . It will be used in Remark 37 to prove the subtlety of characterizing strong normalization with finiteness spaces. The Turing fixed-point combinator Θ has been used in the Introduction to construct (Θ) λx.(x + y) as an example of non-deterministic λ-term morally reducing to normal forms with infinite coefficients (Eq. (3)). Notice that, by abstraction linearity, (Θ) λx.(x+y) = (Θ) (λx.x+λx.y), however (Θ) (λx.x + λx.y) = (Θ) λx.x + (Θ) λx.y, because application is not linear in the argument. This distinction is crucial: the latter term reduces to ((Θ) λx.x) + y, with (Θ) λx.x reducing to itself without producing any further occurence of y. 2.2

Resource Calculus Δ and Taylor Expansion

The syntax. Resource terms and bags are given by mutual induction: resource terms Δ :

s : = x | λx.s | s s

bags Δ ! :

s := 1 | s · s

subject to both α-equivalence and permutativity of (·): we most often write [s1 , . . . , sn ] for s1 · (· · · · (sn · 1) · · · ) and then [s1 , . . . , sn ] = [sσ(1) , . . . , sσ(n) ] for any permutation σ. In other words, bags are finite multisets of terms. Linear Extension. Let R be a rig (a.k.a. semi-ring). Of particular interest are the rigs B : = ({0, 1} , max, min) of booleans, N : = (N, +, ×) of non-negative

412

M. Pagani et al. Table 1. Definition of the Taylor expansion ( )∗ and of its support T ( )

integers and Q : = (Q, +, ×) of rational numbers. We denote as R[[Δ]] (resp. R[[Δ ! ]]) the set of all formal (finite or infinite) linear combinations of resource terms (resp. bags) with coefficients in R. If a ∈ R[[Δ]] (resp. a ∈ R[[Δ ! ]]) and s ∈ Δ (resp. s ∈ Δ ! ), then as ∈ R (resp. as ∈ R) denotes the coefficient of s in a (resp. s in a). As is well-known, R[[Δ]] is endowed with a structure of Rmodule, where addition and scalar multiplication are defined component-wise, i.e. for a, b ∈ R[[Δ]] and α ∈ R: (a + b)s : = as + bs , and (αa)s : = αas . We will write |a| ⊆ Δ for the support of a: |a| = {s ∈ Δ; as = 0}. Moreover, each resource calculus constructor extends to R[[Δ]] componentwise, i.e. for any a, a1 , . . . , an ∈ R[[Δ]] and a ∈ R[[Δ ! ]], we set:   a a : = s∈Δ, s∈Δ ! as as s s, λx.a : = s∈Δ as λx.s, [a1 , . . . , an ] : =

 s1 ,...,sn ∈Δ

(a1 )s1 · · · (an )sn [s1 , . . . , sn ] .

Notice that the last formula is coherent with the notation of the concatenation of bags as a product since it expresses a distributivity law. In particular, we  1 n a . denote an : = [a, . . . , a] and if Q ⊆ R, a! : = n!    n∈N

n

In the case R = B, notice that B[[Δ]] is the power-set lattice P (Δ), so that we can use the set-theoretical notation: e.g. writing s ∈ a instead of as = 0, or a ∪ b for a + b. Also, in that case the preceding formulas lead to: for a, b ⊆ Δ, ∈ a}, a! : = {[s1 , · · · , sn ] ; s1 , · · · , sn ∈ a}, t ∈ Δ, a ⊆ Δ ! : λx.a : = {λx.s; s  t a : = {t s; s ∈ a} and a a : = {s a; s ∈ a}. Taylor expansion. Ehrhard and Regnier have used in [7] the rig of rational numbers to express the λ-terms as formal combinations in Q[[Δ]]. We refer to ∗ this translation ( ) : Λ+ → Q[[Δ]] as the Taylor expansion and we recall it in Table 1a by structural induction. The supports of these expansions can be seen as a map T ( ) : Λ+ → B[[Δ]] and directly defined by induction as in Table 1b. Example 2. From the above definitions we have: Δ∗ =

 1 λx. x xn , n! n

Δ3 ∗ =



1 λx. x xn  xm . n!m! n,m

Strong Normalizability as a Finiteness Structure

413

Let us denote as δn (resp. δn,m ) the term λx. x xn (resp. λx. x xn  xm ). We can then write:  1  1 δn0  [δn1 , . . . , δnk ] , (4) Ω∗ = k! n ,...,n n0 ! · · · nk ! k 0 k  1  1 δn0 ,m0  [δn1 ,m1 , . . . , δnk ,mk ] . (5) Ω3 ∗ = k! n0 ,...,nk n0 !m0 ! · · · nk !mk ! k

m0 ,...,mk

It is clear that the resource terms appearing with non-zero coefficients in M ∗ describe the structure of M taking an explicit number of times the argument of each application, and this recursively. The rˆ ole of the rational coefficients will be clearer once defined the reduction rules over Δ (see Example 5). Operational Semantics. Let us write degx (t) for the number of free occurrences of a variable x in a resource term t. We define the differential substitution of a variable x with a bag [s1 , . . . , sn ] in a resource term t, denoted ∂x t · [s1 , . . . , sn ]: it is a finite formal sum of resource terms, which is zero whenever degx (t) = n; otherwise it is the sum of all possible terms obtained by linearly replacing each free occurrence of x with exactly one si , for i = 1, . . . , n. Formally, ⎧

t sf (1) /x1 , . . . , sf (n) /xn if degx (t) = n, ⎨ ∂x t · [s1 , . . . , sn ] : = f ∈Sn (6) ⎩ 0 otherwise, where Sn is the group of permutations over n = {1, . . . , n} and x1 , . . . , xn is any enumeration of the free occurrences of x in t, so that t sf (i) /xi denotes the term obtained from t by replacing the i-th free occurrence of x with sf (i) . Then, we give a linear extension  of the differential substitution: if a ∈ R[[Δ]] and a ∈ R[[Δ ! ]], we set: ∂x a · a = t∈Δ,s∈Δ ! at as ∂x t · s. The resource reduction →r is then the smallest relation satisfying: λx.t [s1 , . . . , sn ] →r ∂x t · [s1 , . . . , sn ] and moreover linear and compatible with the resource calculus constructors. Spelling out these two last conditions: for any t, u ∈ Δ, s ∈ Δ ! , a, b ∈ R[[Δ]], α ∈ R \ {0}, whenever t →r a, we have: (compatibility) λx.t →r λx.a, t s →r a s, u t · s →r u a · s, and (linearity) αt + b →r αa + b. Proposition 3 ([7]). Resource reduction →r is confluent over the whole R[[Δ]] and it is normalizing on the sums in R[[Δ]] having a finite support. Proposition 3 shows that any single resource term t ∈ Δ has a unique normal form that we can denote as NF(t). What about possibly infinite linear combinations in R[[Δ]]? We would like to extend the normal form operator componentwise, as follows:  at · NF(t) (7) NF(a) : = t∈Δ

414

M. Pagani et al.

Example 4. The sum in Eq. (7) can be undefined for general a ∈ R[[Δ]]. Take a = λx.x + λx.x [λx.x] + λx.x [λx.x [λx.x]] + . . . : any single term of this sum reduces to λx.x, hence NF(a)λx.x is infinite. Another example is given by the Taylor  term in Eq. (3):  expansion of ∗the one can check that the sum defining NF ((Θ) λx.(x + y)) y following Eq. (7) is infinite because, for all n ∈ N, there is a resource term of the form tn  [(λx.x)n , λx.y] ∈ T ((Θ) λx.(x + y)), reducing to y. A closer inspection of the resulting coefficients (that we do not develop here) moreover shows that this infinite sum has unbounded partial sums in Q, hence it diverges in general. In fact, Corollary 39 ensures that if a is the Taylor expansion of a strongly normalizing non-deterministic λ-term then Eq. (7) is well-defined. ∗

Example 5. Recall the expansions

of Example 2 and consider ((Δ) λx.x) =  1 n k . The resource reduction applied to a term λx. x [x ] (λx.x) n,k n!k! of this sum gives zero except for k = n + 1; in this case we get  latter ∗ 1 λx.x [(λx.x)n ], (n+1)! λx.x [(λx.x)n ]. Hence we have: ((Δ) λx.x) →r n n! because the coefficient (n + 1)! generated by the reduction step is erased by the 1 in the definition of Taylor expansion. Then, the term λx.x [(λx.x)n ] fraction k! reduces to zero but for n = 1, in the latter case giving λx.x. So we have: ∗ ∗ NF(((Δ) λx.x) ) = λx.x = (NF((Δ) λx.x)) . The commutation between computing normal forms and Taylor expansions has been proven in general for deterministic λ-terms [6]4 and witnesses the solidity of the definitions. The general case for Λ+ is still an open issue. Example 6. Recall the notation of Eq. (4) from Example 2 expressing the sum Ω∗ . All terms with n0 = k + 1 reduce  in one  step. For n0 = k + 1, we  to zero have that a single term rewrites to f ∈Sk δnf (1) δnf (2) , . . . , δnf (k) , which is a sum of terms still in T (Ω), but with smaller size w.r.t. the redex. Therefore, every term of Ω∗ eventually reduces to zero, after a reduction sequence whose length depends on the initial size of the term, and whose elements are sums with supports in T (Ω). This is in some sense the way Taylor expansion models the unbounded resource consumption of the loop Ω →β Ω in λ-calculus. We postpone the discussion of the reduction of Ω3 ∗ until Remark 37. 2.3

Finiteness Structures Induced by Antireduction

Let us get back to Eq. (7), andconsider it pointwise: for all s ∈ Δ in normal form, we want to set NF(a)s = t∈Δ at · NF(t)s . Notice that this series can be obtained as the inner product between a and the vector ↑s with (↑s )t = NF(t)s : one can think of s as a test, that a passes whenever the sum converges. There is one very simple condition that one can impose on a formal series to ensure its convergence: just assume there are finitely many non-zero terms. This seemingly dull remark is in fact the starting point of the definition of finiteness 4

The statement proven in [6] is actually more general, because it considers (possibly infinite) B¨ ohm trees instead of the normal forms.

Strong Normalizability as a Finiteness Structure

415

spaces, introduced by Ehrhard [3] and discussed in the Introduction. The basic construction is that of a finiteness structure: Definition 7 ([3]). Let A be a fixed set. A structure on A is any set of subsets A ⊆ P (A). For all subsets a and a ⊆ A, we write a ⊥ a whenever a∩a is finite. For all structure A ⊆ P (A), we define its dual A⊥ = {a ⊆ A; ∀a ∈ A, a ⊥ a }. A finiteness structure on A is any such A⊥ . ⊥

Notice that: A ⊆ A⊥⊥ , also A ⊆ A entails A ⊆ A⊥ , hence A⊥ = A⊥⊥⊥ . Let C0 = {|↑s | ; s ∈ Δ, in normal form} ⊆ P (Δ), we obtain that |a| ∈ C0 ⊥ iff Eq. (7) involves only pointwise finite sums. So, one is led to focus on support sets only, leaving out coefficients entirely. Henceforth, unless specified otherwise, we will thus stick to the case of R = B, and use set-theoretical notations only. This approach of ensuring the normalization of Taylor expansion via a finiteness structure was first used by Ehrhard [4] for a non-deterministic variant of System F. Our paper strengthens Ehrhard’s result in several directions. In order to state them, we introduce a construction of finiteness structures on Δ induced by anticones for the reduction order defined as follows: Definition 8. For all s, t ∈ Δ, we write t ≥ s whenever there exists a reduction t →∗r a with s ∈ a. It should be clear that this defines a partial order relation (in particular we have antisymmetry because →r terminates). Definition 9. If a ⊆ Δ, ↑a : = {t ∈ Δ; ∃s ∈ a, t ≥ s} is the cone of antireduction over a.5 For any structure T ⊆ P (Δ), we write  (T) = {↑a ; a ∈ T}. We can consider the elements of T as tests, and say a set a ⊆ Δ passes a test ⊥ a ∈ T iff a ⊥ ↑a  . The structure of sets that pass all tests is exactly  (T) . Then Ehrhard’s result can be rephrased as follows: ⊥

Theorem 10 ([4]). If M ∈ Λ+ is typable in System F then T (M ) ∈  (Ssgl ) where Ssgl : = {{s} ; s ∈ Δ}.

Notice that, in contrast with the definition of C0 , Ssgl is in fact not restricted to normal forms. Our paper extends this theorem in three directions: first, one can relax the condition that M is typed in System F and require only that M is strongly normalizable; second, the same result can be established for sets of “tests” larger (hence more demanding) than Ssgl ; third, the implication can be reversed for a suitable set of tests T, i.e. M is strongly normalizable iff ⊥ T (M ) ∈  (T) (and we do need T to provide more tests than just singletons: see Remark 37). In order to state our results precisely, we need to introduce a few more notions. Definition 11. We say that a resource term s is linear whenever each bag appearing in s has cardinality 1. A set a of resource terms is said linear whenever all its elements are linear. We say a is bounded whenever there exists a number n ∈ N bounding the cardinality of all bags in all terms in a. We then write 5

 Observe that ⏐{s} = ↑s (up to the identification of B[[Δ]] with P (Δ)).

416

M. Pagani et al. Table 2. The intersection type assignment system D+ for Λ+

L : = {a ⊆ Δ; a linear}

and

B : = {a ⊆ Δ; a bounded} .

We also denote as (M ) the subset of the linear resource terms in T (M ). Notice that (M ) is always non-empty and can be directly defined by replacing the definition of T ((M ) N ) in Table 1b with: ((M ) N ) : = {s [t] ; s ∈ (M ) , t ∈ (N )}. Observe that Ssgl , L ⊆ B. We can sum up our results Corollary 30 and Theorem 36 as follows: Theorem 12. Let M ∈ Λ+ : ⊥

– If M is strongly normalizing, then T (M ) ∈  (T) as soon as T ⊆ B. ⊥ – If T (M ) ∈  (T) with L ⊆ T, then M is strongly normalizing.

3

Strongly Normalizing Terms Are D+ Typable

Intersection types are well known, as well as their relation with normalizability. We refer to [2] for the original system with subtyping characterizing the set of strongly normalizing λ-terms, and [1,8] for simpler systems. However, as far as we know, the literature about intersection types for non-deterministic λ-calculi is less well established and in fact we could find no previous characterization of strong normalization in a non-deterministic setting. Hence, we give in Table 2 a variant of Krivine’s system D [8], characterizing the set of strongly normalizing terms in Λ+ . In this section, we only prove that strongly normalizing terms are typable (Theorem 14): the reverse implication follows from the rest of the paper (see Fig. 1). These techniques are standard.

Strong Normalizability as a Finiteness Structure

417

Remark 13. Krivine’s original System D does not have () and (+), but it has the two usual elimination rules for intersection (here derivable). The rule (+) is necessary to account for non-determinism, however adding just (+) to System D is misbehaving. We can find terms M and N , and a context Γ such that (M ) N is typable in System D with (+) under the context Γ but M is not: take Γ = x : A → B ∩B  , y : A → B ∩B  , z : A, observe that (x + y) z = (x) z +(y) z, and thus Γ  (x + y) z : B but x + y is not typable in Γ . This is the reason why we introduce subtyping. Theorem 14. For all M ∈ Λ+ , if M is strongly normalizable, then there exists a derivable judgement Γ  M : A in system D+ . Proof (Sketch). For M a strongly normalizable term, let l (M ) be the maximum length of a reduction from M , and s (M ) the number of symbols occurring in M . By well-founded induction on the pair (l (M ) , s (M )) we prove that there exists nM ∈ N such that for all type B and all n ≥ nM , there is a context Γ and a sequence (A1 , . . . , An ) of types such that Γ  M : A1 → · · · → An → B. The proof splits depending on the structure of M . In case M = M1 + M2 , we apply the induction hypothesis on both M1 and M2 and conclude by rule () and a contravariance property: Γ  M : A whenever Γ   M : A and Γ  Γ  . In case of head-redexes, i.e. M = ((λx.N ) P ) M1 · · · Mq , we apply the induction hypothesis on M  = (N [P/x]) M1 · · · Mq and on P . Then, we conclude via a subject expansion lemma stating that: Γ  (λx.N ) P M1 · · · Mn : A, whenever Γ  (N [P/x]) M1 · · · Mn : A and there exists B such that Γ  P : B. The other cases are similar to the first one.  

4

D+ Typable Terms Are Finitary

This section proves Corollary 30, giving sufficient conditions (to be dispersed, hereditary and expandable, see resp. Definitions 23, 26 and 27) over a structure T in order to have all cones ↑a for a ∈ T dual to the Taylor expansion of any strongly normalizing non-deterministic λ-term. It is easy to check that these conditions are satisfied by the structures B of bounded sets and L of sets of linear terms (Definition 11). Moreover, as an immediate corollary one gets also that any subset T ⊆ B is also such that ⊥ T (M ) ∈  (T) for any strongly normalizable term M ∈ Λ+ , so getting the first Item of Theorem 12. Thanks to previous Theorem 14 we can prove Corollary 30 by a realizability technique on the intersection type system D+ . For a fixed structure S, we associate with any type A a realizer AS (Definition 18). In the case S is adapted (Definition 17), we can prove that AS contains the Taylor expansion of any term of type A and that it is contained in S (Theorem 21). These definitions and theorem are adapted from Krivine’s proof for System D [8]. The crucial point is then to find structures S which are adapted : here is our ⊥ contribution. The structures that we study have the shape  (T) , so that we are speaking of the interaction with cones of anti-reducts of tests in a structure T.

418

M. Pagani et al.

Intuitively, T is a set of tests that can be passed by any term typable in System D+ (hence by any strongly normalizing term). We prove (Lemma 29) that for a structure T, being dispersed, hereditary and expandable is sufficient to guarantee ⊥ that the dual structure  (T) is adapted, then achieving Corollary 30.  Definition 15 (Functional). Given two structures S,  S ⊆ P (Δ), we define  the structure S → S : = f ⊆ Δ; ∀a ∈ S, f  a! ∈ S .   Definition 16 (Saturation).  Let S, S ⊆ P (Δ). We say S is S-saturated if  ! ! !  ∀e, f0 , . . . , fn ∈ S, ∂x e · f0 f1 . . . fn ∈ S implies λx.e f0 ! f1 ! . . . fn ! ∈ S .

Definition 17 (Adaptedness). For all S ⊆ P (Δ) we   set Sh : = {{x} ; x ∈ V} ∪ x a1 ! · · · an ! ; x ∈ V, n > 0 and ∀i, ai ∈ S and say S is adapted if: 1. S is S-saturated; 2. Sh ⊂ (S → Sh ) ⊂ (Sh → S) ⊂ S; 3. S is closed under finite unions: ∀b, b ∈ S, b ∪ b ∈ S. Definition 18 (Realizers). Let S ⊆ P (Δ). To each type A of System D+ , we associate a structure AS defined inductively (X being a propositional variable): XS : = S, A → BS : = AS → BS , A ∩ BS : = AS ∩ BS . Lemma 19. Let S be an adapted structure, then for every type A, AS is S-saturated, closed under finite unions and Sh ⊆ AS ⊆ S. Lemma 20 (Adequacy). If S ⊆ P (Δ) is adapted, x1 : A1 , . . . , xn : An  M : B and for all 1 ≤ i ≤ n, ai ∈ Ai S , then ∂x1 ,...,xn T (M ) · a1 ! , . . . , an ! ∈ BS . Proof (Sketch). By structural induction on the derivation of x1 : A1 , . . . , xn : An  M : B. The cases where the last rule is (→i ) or (+) use respectively the facts that the realizers are saturated and closed by finite unions. The case where the last rule is () is an immediate consequence of the induction hypothesis and   of a lemma stating that A  B implies AS ⊆ BS . Theorem 21. If S ⊆ P (Δ) is adapted and M is typable in System D+ , then T (M ) ∈ S. Proof. Let Γ  M : B. For any x : A in Γ , {x} ∈ Sh ⊂ AS by Lemma 19 ! ! and Definition 17. Hence, T (M ) = ∂x1 ,...,xn T (M ) · {x1 } , . . . , {xn } ∈ BS by   Lemma 20. Again by Lemma 19, BS ⊆ S, so T (M ) ∈ S. Now we look for conditions to ensure that a structure S is adapted. These conditions (Definitions 23, 26 and 27) are quite technical but they are easy to check, in particular the structures B and L enjoy them (Remark 28). Definition 22. The height h (s) of a resource term s is defined inductively: h (x) : = 1, h (λx.s) : = 1 + h (s) and h (s0  [s1 , . . . , sn ]) : = 1 + maxi (h (si )).

Strong Normalizability as a Finiteness Structure

419

Definition 23 (Dispersed). A set a ⊆ Δ is dispersed whenever for all n ∈ N, and all finite set V of variables, the set {s ∈ a; h (s) ≤ n and fv (s) ⊆ V } is finite. A structure S is dispersed whenever ∀a ∈ S, a is dispersed. Definition 24. Let s ∈ Δ and x ∈ fv (s). We immediate subterm pro  define ! jections πx s ∈ P (Δ), π0 s ∈ P (Δ), π 1 s ∈ P Δ and π1 s ∈ P (Δ) as follows: – if s = λx.t then πx s = {t}; otherwise πx s = ∅; – if s = t u then π0 s = {t}, π 1 s = {u}, π1 s = |u|; otherwise π0 s = π 1 s = π1 s = ∅. Observe that up to α-conversion and the hypothesis x ∈ fv (s), the abstraction case is exhaustive. These functions obviously extend to sets of terms, up to some care about free variables. If V ⊆ V is a set of variables, we write ΔV for the set of resource λ-terms with free variables in V . Definition 25. For all V ⊆ V and a ⊆ ΔV , let   ! π0 s ⊆ Δ V , π1 a : = π 1 s ⊆ ΔV , π0 a : = s∈a

π1 a : =

s∈a

and if moreover x ∈ V , then let πx a : =



π1 s ⊆ Δ V ,

s∈a

 s∈a

πx s ⊆ ΔV ∪{x} .

Definition 26 (Hereditary). A structure S ⊆ P (Δ) is said to be hereditary if, S is downwards closed, and for all a ∈ S, π0 a ∈ S, π1 a ∈ S and for all x ∈ V \ fv (a), πx a ∈ S. Definition 27 (Expandable). A structure S ⊆ P (Δ) is said to be expandable if, for all x ∈ V and all a ∈ S, we have {s [x] ; s ∈ a} ∈ S. Remark 28. The structures Ssgl (Theorem 10) and L, B (Definition 11) are dispersed and expandable. The last two are also hereditary, while Ssgl is not. Lemma 29. For any structure T which is dispersed, hereditary and expandable, ⊥ we have that  (T) is adapted. Proof (Sketch). One has to prove the three conditions of Definition 17. Heredity and dispersion are used to obtain saturation (Condition 1). The inclusions of ⊥ Condition 2 need all hypotheses and an auxiliary lemma proving that  (T) h ⊆ ⊥  (T) , for which dispersion is crucial. Finally, the closure under finite unions (Condition 3) is satisfied by all finiteness structures.   Corollary 30. Let T ⊆ P (Δ) be dispersed, hereditary and expandable. For ⊥ every strongly normalizable term M , we have T (M ) ∈  (T) . Hence, this holds for T ∈ {L, B} and for any of their subsets, such as Ssgl ⊂ B. Proof. The general statement follows from Theorems 14 and 21 and Lemma 29. ⊥ Remark 28 implies T (M ) ∈  (T) for T ∈ {B, L}. The rest of the statement ⊥   follows because, for any structures S, S , S ⊆ S implies S ⊆ S⊥ .

420

5

M. Pagani et al.

Finitary Terms Are Strongly Normalizing

In this section we prove Theorem 36, giving a sufficient condition for a structure T to be able to test strong normalization. The condition is that T includes at least L, i.e. the set of all sets of linear terms6 . The proof is by contraposition, suppose that M is divergent, then T (M ) is not dual to some cone ↑a , with a ∈ L. The proof enlightens two kinds of divergence in λ-calculus: the one generated by looping terms: Ω →β Ω →β . . . and the other generated by infinite reduction sequences (Mi )i∈N with an infinite number of different terms: Ω3 →β (Ω3 ) Δ3 →β ((Ω3 ) Δ3 ) Δ3 →β . . . (see Example 1).  In the first case, the cone ⏐ (Ω) of the linear expansion (Definition  11) of the looping term Ω suffices to show up the divergence, since T (Ω) ∩ ⏐ (Ω) is infinite. Indeed the Taylor expansion of a looping term, say T (Ω), is a kind of “contractible space”, where any resource term reduces to a smaller term within the same Taylor expansion or vanishes (see Example 6). In particular, there are unboundedly large resource terms reducing to the linear expansion (Ω). In the case of an infinite reduction sequence of different terms, one should take, basically, the cone of all linear expansions of the terms occurring in the sequence: the linear expansion of a single term (or of a finite set of terms) might not suffice to test this kind of divergence. For example, T (Ω3 )∩ ⏐ (Ω3 ) is finite,  ⊥ while T (Ω3 ) ∩ ⏐{ (Ω3 ), ((Ω3 ) Δ3 ), . . . } is infinite, so T (Ω3 ) ∈ /  (L) . In the presence of the non-deterministic sum +, we have a third kind of divergence, which is given by infinite reduction sequences of terms (Mi )i∈N which are pairwise different but whose Taylor expansion support repeats infinitely many times: consider, e.g., the reducts of (Θ) (λx.x + y). We prove that this kind of divergence is much more similar to a loop rather than to a sequence of different λ-terms. In particular, there is a single linear resource term (depending on the reduction sequence) whose cone is able to show up the divergence. Indeed, most of the effort in the proof of Theorem 36 is devoted to deal with this kind of “looping Taylor expansion”. Namely, Definition 31 gives a notion of non-deterministic reduction  allowing Lemma 35, which is the key statement used in the proof of Theorem 36 for dealing with both the divergence of looping terms (like Ω) and that of looping Taylor expansions (like (Θ) (λx.x + y)). We introduce a reduction rule  on Λ+ which corresponds to one step of β-reduction and a potential loss of some addenda in a term. For that, we need an order  on Λ+ expressing this loss. For instance, (Θ) (λx.x + y) + y  (Θ) λx.x, thus (Θ) (λx.x + y) ∗ (Θ) λx.x, and similarly, (Θ) (λx.x + y) ∗ y. Definition 31 (Partial reduction). We write M  N if there exists P such that M →β P and P  N , where the partial order  over Λ+ is defined as the least order such that M  M + N ; N  M + N and if M  N then: M + P  N + P , λx.M  λx.N , (M ) P  (N ) P , and (P ) M  (P ) N . 6

This condition can be slightly weakened replacing L with: {a ⊆ Δ; a linear and fv (a) finite}. However, we prefer to stick to the more intuitive definition of L.

Strong Normalizability as a Finiteness Structure

421

A reduction M  N is at top level if M = (λx.M  ) M  → M  [M  /x]  N . Write s > t whenever s ≥ t (Definition 8) and s = t: this is a strict partial order relation. Lemma 32. If M  N and t ∈ T (N ), then there exists s ∈ T (M ) such that s ≥ t. If moreover, M  N is at top level, then s > t. Lemma 33. Let M  N and u ∈ Δ. If T (N ) ∩ ↑u is infinite, then T (M ) ∩ ↑u is also infinite. Definition 34. The height h (M ) of a term M ∈ Λ+ is defined inductively as follows: h (x) : = 1, h (λx.M ) : = 1 + h (M ), h ((M ) N ) : = 1 + max(h (M ) , h (N )) and h (M + N ) : = max(h (M ) , h (N )). Lemma 35. Let (Mi )i∈N be a sequence. If ∀i ∈ N, Mi  Mi+1 and (h (Mi ))i∈N is bounded, then there exists a linear term t such that T (M0 ) ∩ ↑t is infinite. Proof (Sketch). First, it is sufficient to address the case of a sequence (Si )i∈N of simple terms (i.e. without + as the top-level constructor) such that Si  Si+1 for all i ∈ N and (h (Si ))i∈N is bounded. Besides, by Lemma 33, it is sufficient to have T (Si0 ) ∩ ↑t infinite for some i0 ∈ N. Then, by induction on h = max {h (Si ) ; i ∈ N}, we show that there exists N i0 ∈ N and a sequence (sj )j∈N ∈ T (Si0 ) such that s0 is linear and, for all j ∈ N, sj+1 > sj . Since > is a strict order relation, this implies that the set {sj ; j ∈ N} ⊆ T (Si0 ) ∩ ↑s0 is infinite. First assume that there are infinitely many top level reductions. Observe that, since h (Si ) ≤ h and fv (Si ) ⊆ fv (S0 ) for all i, the set {T (Si ) ; i ∈ N} is finite. Hence there exists an index i0 ∈ N such that {i ∈ N; T (Si ) = T (Si0 )} is infinite. As there are infinitely many top level reductions, there are i1 and i2 such that i0 < i1 < i2 , the reduction i1 is at top level and T (Si2 ) = T (Si0 ). We inductively define the required sequence by choosing arbitrary s0 ∈ (Si0 ) ⊆ T (Si0 ), and by iterating Lemma 32: for sj ∈ T (Si0 ) = T (Si2 ), we obtain sj+1 ∈ T (Si0 ) with sj+1 > sj since the reduction i1 is at top level. Now assume that there are only finitely many top level reductions. Let i1 be such that no reduction Si  Si+1 with i ≥ i1 is at top level. Either for all  , and we conclude by applying the induci ≥ i1 , Si = λx.Mi with Mi  Mi+1  ) ; or for all i ≥ i1 , Si = (Mi ) Ni so tion hypothesis to the sequence (Mi+i 1 i∈N   that (Mi )i≥i1 and (Ni )i≥i1 are sequences of terms, at least one of them involving infinitely many partial reductions. In this case, assume for instance that we can  )i∈N of partial reductions. It extract from (Mi )i≥i1 an infinite subsequence (Mφ(i)    N provides i0 and a sequence (sj ) ∈ T (Mφ(i ) ) with s0 linear and, sj+1 > sj . Fix 0     [t] for all j ∈ N.   t ∈ (Nφ(i  ) ) arbitrarily. So we set i0 = φ(i0 ) and sj = sj 0

Theorem 36. Let T be a structure such that L ⊆ T. If T (M ) ∈  (T) M is strongly normalizable. In particular, this holds for T ∈ {L, B}.



then

422

M. Pagani et al.

Proof. Assume that (Mi )i∈N is such that M = M0 and for all i, Mi →β Mi+1 . ⊥ We prove that T (M ) ∈  (T) by exhibiting a ∈ T such that T (M ) ⊥ ↑a . If (h (Mi ))i∈N is bounded, then fix a = {t} with t given by Lemma 35. Otherwise, ∀i ∈ N, fix ti ∈ (Mi ) such that h (ti ) = h (Mi ). Lemma 32 implies that there is si ∈ T (M ) such that si ≥ ti . Denote by s (s) the number of symbols occurring in s. Since there is no duplication in reduction →r it should be clear that if s ≥ t then s (s) ≥ s (t). Besides, s (s) ≥ h (s). Therefore, since {h (Mi ) ; i ∈ N} is unbounded,  {h (ti ) ; i ∈ N}, {s (ti ) ; i ∈ N} and {s (si ) ; i ∈ N} are unbounded. Fix a = i∈N (Mi ) ∈ T, we have proved that T (M ) ∩ ↑a is infinite.   Notice that the structure Ssgl of singletons used in [4] does not enjoy the hypothesis of Theorem 36 (L ⊂ Ssgl ). In fact: ⊥

Remark 37. We prove that T (Ω3 ) ∈  (Ssgl ) , although Ω3 is not normalizing. Recall from Example 2, that the support of the Taylor expansion of Ω3 is made of terms of the form δn0 ,m0  [δn1 ,m1 , . . . , δnk ,mk ] (for k, ni , mi ∈ N). Write Δh = {δ−,−  [. . . δ−,− . . . ] · · · [. . . δ−,− . . . ] with h bags}: in particular T (Ω3 ) = Δ1 . One can easily check that if s ∈ Δh and s ≥ s , then s ∈ Δh with h ≤ h . A careful inspection of such reductions shows that they are moreover reversible: for all s ∈ Δh and all h ≤ h there is exactly one s ∈ Δh such that s ≥ s . It follows ⊥ that Δ1 ∩ ↑ s is either empty or a singleton. Therefore T (Ω3 ) ∈  (Ssgl ) .

6

Conclusion

We achieved all implications of Fig. 1, but the rightmost one, concerning the finiteness of the coefficients in the normal form of the Taylor expansion of a strongly normalizing λ-term (recall Eq. (2) in the Introduction). Thanks to the definition of cones (Definition 9) we immediately have the following lemma, which is the last step to Corollary 39.  ⊥ T, Lemma 38. Let T be a structure. If T (M ) ∈  (T) , then ∀t ∈ NF(T (M ))t is finite. Applying Corollary 30 and Lemma 38 to a structure like B or Ssgl , we get: Corollary 39. Given a non-deterministic λ-term M , if M is strongly normalizable, then NF(T (M ))t is finite for all t ∈ Δ.

References 1. van Bakel, S.: Complete restrictions of the intersection type discipline. Theor. Comput. Sci. 102(1), 135–163 (1992). http://dx.doi.org/10.1016/0304-3975(92)90297-S 2. Barendregt, H., Coppo, M., Dezani-Ciancaglini, M.: A filter lambda model and the completeness of type assignment. JSL 48, 931–940 (1983) 3. Ehrhard, T.: Finiteness spaces. Math. Struct. Comput. Sci. 15(04), 615–646 (2005)

Strong Normalizability as a Finiteness Structure

423

4. Ehrhard, T.: A finiteness structure on resource terms. In: Proceedings of the 25th Annual IEEE Symposium on Logic in Computer Science, LICS 2010, 11–14 July 2010, Edinburgh, United Kingdom, pp. 402–410. IEEE Computer Society (2010). http://dx.doi.org/10.1109/LICS.2010.38 5. Ehrhard, T., Regnier, L.: The differential lambda-calculus. Theor. Comput. Sci. 309(1), 1–41 (2003) 6. Ehrhard, T., Regnier, L.: B¨ ohm trees, Krivine’s Machine and the Taylor expansion of lambda-terms. In: Beckmann, A., Berger, U., L¨ owe, B., Tucker, J.V. (eds.) CiE 2006. LNCS, vol. 3988, pp. 186–197. Springer, Heidelberg (2006) 7. Ehrhard, T., Regnier, L.: Uniformity and the Taylor expansion of ordinary lambdaterms. Theor. Comput. Sci. 403(2–3), 347–372 (2008) 8. Krivine, J.L.: Lambda-Calculus, Types and Models. Ellis Horwood, Chichester (1993) 9. Laird, J., Manzonetto, G., McCusker, G., Pagani, M.: Weighted relational models of typed lambda-calculi. In: Proceedings of 28th Annual ACM/IEEE Symposium on Logic in Computer Science (LICS 2013), pp. 25–28, New Orleans, USA, pp. 301–310, June 2013 10. Lamarche, F.: Quantitative domains and infinitary algebras. Theor. Comput. Sci. 1, 37–62 (1992) 11. Pagani, M.: The cut-elimination theorem for differential nets with promotion. In: Curien, P.-L. (ed.) TLCA 2009. LNCS, vol. 5608, pp. 219–233. Springer, Heidelberg (2009) 12. Pagani, M.: Visible acyclic differential nets, Part I: Semantics. Ann. Pure Appl. Logic 163(3), 238–265 (2012)

Reasoning About Call-by-need by Means of Types Delia Kesner(B) Universit´e Paris-Diderot, SPC, IRIF, CNRS, Paris, France [email protected]

Abstract. We first develop a (semantical) characterization of call-byneed normalization by means of typability, i.e. we show that a term is normalizing in call-by-need if and only if it is typable in a suitable system with non-idempotent intersection types. This first result is used to derive a new completeness proof of call-by-need w.r.t. call-by-name. Concretely, we show that call-by-need and call-by-name are observationally equivalent, so that in particular, the former turns out to be a correct implementation of the latter.

1

Introduction

There is a real gap between the well-known operational semantics of the callby-name λ-calculus and the actual implementations of lazy functional languages such as Miranda or Haskell. Indeed, call-by-name re-evaluates an argument each time it is used – an operation which is quite expensive – while lazy languages store the value of an argument the first time it is evaluated, thus avoiding the need for any subsequent re-evaluations. For example, consider the term t0 = Δ(II), where Δ = λx.xx and I = λz.z. Call-by-name duplicates the argument II, while lazy languages first reduces II to the value I so that further uses of this argument do not need any more to evaluate it again. However, lazy languages should not be confused with call-by-value: values are only consumed in the former when they are needed/required, thus completely ignoring the other arguments. E.g. if t1 = (λz.I)(II), the argument II is evaluated in call-by-value to produce the final answer I, whereas this evaluation does not take place in any lazy language (see e.g. [2, 3, 9, 13]). Said differently, lazy languages do not compute unneeded terms, and terms should not be duplicated unless and until they have been reduced to values. According to the previous paragraphs, lazy languages cannot be modeled by call-by-name, nor by call-by-value; they are indeed specified by call-by-need, due to Wadsworth [30], and extensively studied e.g. in [1–3, 9,13, 25]. Despite the fact that the equational theories of call-by-name and call-by-need are not the same [2], call-by-need is observationally equivalent (i.e. sound and complete w.r.t.) to call-by-name. The completeness result, operationally developed e.g. in [2, 25], is quite involved and makes use of different syntactical tools such as commutation diagrams, sharing, residual theory and standardization. c Springer-Verlag Berlin Heidelberg 2016  B. Jacobs and C. L¨ oding (Eds.): FOSSACS 2016, LNCS 9634, pp. 424–441, 2016. DOI: 10.1007/978-3-662-49630-5 25

Reasoning About Call-by-need by Means of Types

425

In this paper we make use of typing systems to provide a semantical argument showing that call-by-need and call-by-name are observationally equivalent. More precisely, we develop a new completeness proof of call-by-need w.r.t. call-by-name which follows from two different implications: every term having a normal form in call-by-name is typable and every typable term has a normal form in call-by-need. The second implication is obtained from a stronger characterization property stating that the set of typable terms and the set of call-by-need normalizing terms are the same. This result is completely new, and quite surprising, and constitutes one of the main contributions of this work. In order to achieve our goal we use intersection types, originally introduced in [11], being able to provide models of the λ-calculus [4], and characterizations of head [10] as well as weakly [8, 10] and strongly normalizing λ-terms [22]. Typically, intersection types are idempotent [4]. However, in the last years, growing interest has been devoted to non-idempotent intersection types [6, 19], since they allow for reasoning about quantitative properties of terms, both from a syntactical and a semantical point of view. Different assignment systems with nonidempotent intersection types have been studied in the literature for different purposes [5,7,14, 15, 18, 20,21, 23, 26,27]. Reasoning about call-by-name and call-by-need is achieved in this paper by introducing two (non-idempotent) typing systems, called V and A respectively. System V (resp. A) characterizes the set of name (resp. need) normalizing terms, i.e. a term t is typable in system V (resp. A) if and only if t is normalizing in call-by-name (resp. call-by-need). While our characterization of call-by-name normalization can be seen as a simple adaptation of the existing one for headnormalization [15], this is, to the best of our knowledge, the first time that call-by-need normalization is characterized by means of semantical/logical tools. Non-idempotent intersection types provide a full characterizations of call-byname/need normalizing terms by using very simple combinatorial arguments, no reducibility technique are used. More precisely, there is a measure, defined on type derivations, which strictly decreases during name/needed reduction. This gives quantitative information about call-by-name/need reduction sequences. The following list summarizes the more important points of our work: 1. We provide a characterization of name-normalization by adapting an existing one for head-normalization. 2. We develop a completely new characterization of need-normalization by means of typability. This characterization, as well as the one for namenormalization, is arithmetical, i.e. no reducibility arguments are needed. 3. Based on the previous points, we show that call-by-need and call-by-name are observationally equivalent, so that in particular, we obtain a new (semantical) completeness proof of call-by-need w.r.t. call-by-name. Related Works: Besides all the aforementioned works that use intersection types to characterize operational properties of λ-calculi, two other works make use of typing systems to reason about lazy languages. In [24], the resource conscious character of linear logic is used to explain call-by-need in the spirit of the CurryHoward isomorphism. In [16] non-idempotent intersection types are related to

426

D. Kesner

needed redexes: a redex r in a term t is needed iff r, or some residual of r, is reduced in every reduction of t to normal form. Despite the use of similar typing techniques to identify needed redexes in well-typed terms, no full characterization of needed normalization is obtained in [16]. Structure of the Paper: Section 2 introduces syntax and operational semantics of call-by-name and call-by-need. Section 3 defines non-idempotent type assignment disciplines for both calculi. Section 4 (resp. Sect. 5) gives a characterization of name (resp. need)-normalizing terms by means of the intersection type systems introduced in Sect. 3. Section 6 presents the completeness proof by using the semantical/logical arguments developed in the previous sections. Section 7 concludes and suggests some further directions of research.

2

Call-by-name and Call-by-need

This section describes the syntax and the operational semantics of our languages. We start by introducing call-by-name in the setting of the lambda-calculus, as done in the seminal work of Plotkin [28]. Given a countable infinite set of symbols x, y, z, . . ., we define terms, values and name contexts as follows: Terms Values Name contexts

t, u ::= x | tu | λx.t v ::= λx.t E ::=  | E t

The set of terms is denoted by Ta . We write I for the identity function λx.x. The sets of free and bound variables of a term t, written respectively fv(t) and bv(t), are defined as usual. We work with the standard notion of α-conversion i.e. renaming of bound variables for abstractions. Substitutions are (finite) functions from variables to terms denoted by {x1 /u1 , . . . , xn /un } (n ≥ 0). Application of the substitution σ to the term t, written tσ, is defined on α-equivalence classes, as usual, so that capture of free variables cannot hold. The call-by-name calculus is given by the set of terms Ta and the reduction relation →name , which is the closure by name contexts E of the rewriting rule (λx.t)u →β t{x/u}. An example of name-reduction sequence is (λf.f I(f I))(λw.(II)w) →name (λw.(II)w)I((λw.(II)w)I) →name (II)I((λw.(II)w)I) 3name (λw.(II)w)I →name (II)I →name II →name I Remark that t →name t implies fv(t) ⊇ fv(t ), and notice that reduction is weak, so that we never reduce under λ-abstractions, which are considered to be values. The call-by-need calculus we use in this paper was introduced in [1]. It is more concise than previous specifications of call-by-need [2, 3, 9,25], but it is operationally equivalent to them, so that the results of this paper could also be presented by using other alternative specifications.

Reasoning About Call-by-need by Means of Types

427

Given a countable infinite set of symbols x, y, z, . . . we define different syntactic categories for terms, values, list contexts, answers and need contexts: Terms Values List contexts Answers Need contexts

t, u ::= x | tu | λx.t | t[x/u] v ::= λx.t L ::=  | L[x/t] A ::= L[λy.t] M, N ::=  | Nt | N[x/t] | N[[x]][x/M]

Terms of the form t[x/u] are closures, and [x/u] is said to be an explicit substitution (ES). The set of all the terms is denoted by Te ; remark that Ta ⊂ Te . The notions of free and bound variables are defined as expected, in particular, fv(t[x/u]) := fv(t)\{x}∪fv(u), fv(λx.t) := fv(t)\{x}, bv(t[x/u]) := bv(t) ∪ {x} ∪ bv(u) and bv(λx.t) := bv(t) ∪ {x}. We use N[t] (resp. L[t]) for the term obtained by replacing the hole of N (resp. L) by the term t. We use the special notation N[[u]] or L[[u]] when the free variables of u are not captured by the context, i.e. there are no abstractions or explicit substitutions in the context that binds the free variables of u. Thus for example, given N = (x)[x/z], we have (yx)[x/z] = N[y] = N[[y]], but (xx)[x/z] = N[x] cannot be written as N[[x]]. Notice the use of this special notation in the last case of needed contexts, an example of such case being (xy)[y/t][x/]. The call-by-need calculus, introduced in [1], is given by the set of terms Te and the reduction relation →need , the union of →dB and →lsv , which are, respectively, the closure by need contexts N of the following rewriting rules: L[λx.t]u →dB L[t[x/u]] N[[x]][x/L[v]] →lsv L[N[[v]][x/v]] An example of need-reduction sequence is (λx1 .I(x1 I))(λy.Iy) →dB x2 [x2 /x1 I][x1 /λy.Iy] →lsv x2 [x2 /(Ix3 )[x3 /I]][x1 /λy.Iy] →dB x2 [x2 /x4 [x4 /I][x3 /I]][x1 /λy.Iy] →lsv I[x2 /I][x4 /I][x3 /I][x1 /λy.Iy]

(I(x1 I))[x1 /λy.Iy] →dB x2 [x2 /(λx3 .Ix3 )I][x1 /λy.Iy] →dB x2 [x2 /x4 [x4 /x3 ][x3 /I]][x1 /λy.Iy] →lsv x2 [x2 /I[x4 /I][x3 /I]][x1 /λy.Iy] →lsv

As for call-by-name, reduction preserves free variables, i.e. t →need t implies fv(t) ⊇ fv(t ). Notice that call-by-need reduction is also weak, so that answers are not need-reducible. Among the free variables of terms we distinguish the set of needed free variables, defined as follows: nv(x)

:= {x}

nv(λx.t) := ∅

:= nv(t)  (nv(t) \ {y}) ∪ nv(u) if y ∈ nv(t) nv(t[y/u]) := nv(t) \ {y} if y ∈ / nv(t)

nv(tu)

A useful property about needed free variables can be stated as follows: Lemma 1. Let t ∈ Te . Then x ∈ nv(t) if and only if there exists N s.t. t = N[[x]].

428

D. Kesner

Thus for example, for t = x1 [x1 /x2 y][x2 /x3 y  ], we have x3 ∈ nv(t) and t = N[[x3 ]], for N = x1 [x1 /x2 y][x2 /y  ]. We now need some general notions applicable to any reduction system R. We denote by R (resp. →+ R ) the reflexive-transitive (resp. transitive) closure of any given reduction relation →R . A term t is in R-normal form, written t ∈ R-nf, if there is no t s.t. t →R t ; and t has an R-normal form iff there is t ∈ R-nf such that t R t . A term t is said to be weakly R-normalizing, or R-normalizing, written t ∈ WN (R), iff t has an R-nf. It is well-known that the set of Ta -terms that are in name-nf can be characterized by the following grammar: Na ::= λx.t | Nva

Nva ::= x | Nva t

Similarly, Te -terms in need-nf can be specified by the following grammar: Ne ::= L[λx.t] | Nae Nae ::= x | Nae t | Nae[x/u] x ∈ / nv(Nae) | Nae[x/Nae] x ∈ nv(Nae) Lemma 2. Let t ∈ Te . Then t ∈ Ne iff t is in need-nf. Proof. The proof proceeds by first showing that t ∈ Nae iff t is in need-nf and t is not an answer. Then the statement of the lemma follows. ♦ In order to relate need-nfs with name-nfs we use the following function defined on Te -terms. x↓ := x ↓ (tu) := t↓ u↓





(λx.t) := λx.t↓ ↓ t[x/u] := t↓ {x/u↓ }

Notice that x ∈ nv(t) implies x ∈ nv(t↓ ). Lemma 3. Let t ∈ Te . Then t in need-nf implies t↓ in name-nf. Proof. We first remark that t{x/u} ∈ Nva iff t ∈ Nva and x ∈ / nv(t), or t and u are in Nva and x ∈ nv(t). The proof then proceeds by showing the following two statements by induction on the grammars. 1) t ∈ Nae implies t↓ ∈ Nva

2) t ∈ Ne implies t↓ ∈ Na

(1) (a). If t = x ∈ Nae, then the statement is straightforward. (b). If t = t1 t2 ∈ Nae, where t1 ∈ Nae, then the i.h. gives t1 ↓ ∈ Nva so that t↓ = t1 ↓ t2 ↓ ∈ Nva. (c). If t = t1 [x/t2 ] ∈ Nae, where t1 ∈ Nae, then the i.h. gives t1 ↓ ∈ Nva. We distinguish two cases. If x ∈ / nv(t1 ), then t↓ = t1 ↓ {x/t2 ↓ } ∈ Nva. If x ∈ nv(t1 ), then necessarily t2 ∈ Nva, so that the i.h. gives t2 ↓ ∈ Nva. Then, t↓ = t1 ↓ {x/t2 ↓ } ∈ Nva. (2) (a). If t = L[λx.u] ∈ Ne, then t↓ is necessarily an abstraction so that t↓ ∈ Na. (b). If t ∈ Ne comes from t ∈ Nae, then we apply the previous point. ♦

Reasoning About Call-by-need by Means of Types

3

429

Non-idempotent Intersection Types

Our results rely on typability of terms in suitable systems with non-idempotent intersection types. This section introduces such type systems. The first one, called V (for values), characterizes the set of name-normalizing terms; it extends the one in [15] which characterizes the set of head normalizing terms in λ-calculus. The second one, called A (for answers), characterizes the set of need-normalizing terms; it extends the one in [18] which characterizes the set of head-linear normalizing terms in λ-calculus with explicit substitutions. In order to introduce the type systems, let us first introduce the notion of type. We denote finite multisets by double curly brackets, so that {{ }} denotes the empty multiset; {{a, a, b}} denotes a multiset having two occurrences of the element a and one occurrence of b. We use for multiset union. Given a constant type a which denotes answers, and a countable infinite set of base types α, β, γ, . . ., we consider types and multiset types defined by the following grammars: Types Multiset types

τ, σ, ρ ::= a | α | M → τ where I is a finite set M ::= {{τi }}i∈I

The type {{ }} plays the rˆ ole of the universal ω type in [8, 10]. Remark that types are strict [12,29], i.e. the right-hand sides of functional types are never multisets. Moreover, they make use of standard notations for multisets, as in [15], so that {{σ, σ, τ }} needs to be understood as the intersection type σ ∧ σ ∧ τ , where the symbol ∧ is commutative and associative but non-idempotent, i.e. σ ∧ σ and σ are not equivalent. Type assignments, written Γ, Δ, are functions from variables to multiset types, assigning the empty multiset to all but a finite set of the variables. The domain of Γ is dom(Γ ) := {x | Γ (x) = {{ }}}. The union of type assignments, written Γ + Δ, is a type assignment defined by (Γ + Δ)(x) := Γ (x) Δ(x), where denotes multiset union; thus, dom(Γ + Δ) = dom(Γ ) ∪ dom(Δ). An example is {x:{{σ}}, y:{{τ }}} + {x:{{σ  }}, z:{{τ  }}} = {x:{{σ, σ  }}, y:{{τ }}, z:{{τ  }}}. For convenience, we write +i∈I Γi for a finite union of type assignments (where I = ∅ gives an empty function), instead of the more traditional notation Σi∈I Γi . When dom(Γ ) and dom(Δ) are disjoint we use Γ ; Δ instead of Γ +Δ, and we write x:{{σi }}i∈I ; Γ , even when I = ∅, for the assignment (x:{{σi }}i∈I ; Γ )(x) = {{σi }}i∈I and (x:{{σi }}i∈I ; Γ )(y) = Γ (y) if y = x. We denote by Γ \\ x the assignment (Γ \\ x)(x) = {{ }} and (Γ \\ x)(y) = Γ (y) if y = x. We write x#Γ iff x ∈ / dom(Γ ). Type judgments have the form Γ t:τ , where Γ is a type assignment, t is a term and τ is a type. The V-type system (V for value) for Ta -terms is given by the typing rules in Fig. 1 and the A-type system (A for answer) for Te -terms is given by the typing rules in Fig. 2. System V can be seen as the restriction of system A to Ta -terms. Given S ∈ {V, A}, a (type) derivation in S is a tree obtained by applying the (inductive) rules of the system S. We write Φ S Γ t:τ if there is a derivation Φ in system S typing t, i.e. ending in the

430

D. Kesner

Fig. 1. The non-idempotent intersection type system V for terms in Ta

Fig. 2. The non-idempotent intersection type system A for terms in Te

type judgment Γ t:τ . A term t is typable in S iff there is a derivation in S typing t. The size of a type derivation Φ is a natural number sz(Φ) denoting the number of nodes of the tree Φ. The constant type a in rules (val) and (ans) is used to type values and answers respectively, and rule (cut) of system A may also type answers (when τ = a and t = L[λx.t ]). The axiom (ax) is relevant (there is no weakening) and the rules (→ e) and (cut) are multiplicative; both characteristics being related to the resource aware semantics of the underlying calculus. A particular case of rule (→ e) (resp. (cut)) is when I = ∅: the subterm u occurring in the typed term tu (resp t[x/u]) turns out to be untyped. Thus for example, if Ω is the (untypable) non-terminating term (λz.zz)(λz.zz), then from the type derivation of x:{{σ}} λy.x:{{ }} → σ we can construct one for x:{{σ}} (λy.x)Ω:σ, and from the type derivation of x:{{σ}} x:σ we can construct one for x:{{σ}} x[y/Ω]:σ. A major difference between system V and the one in [15] is the rule (val) which types any kind of abstraction. Indeed, given Δ = λx.xx, the abstraction λx.ΔΔ –not being head-normalizing in λ-calculus– is typable in our type system (which characterizes name-normalizing terms) but is not typable in [15] (which characterizes head-normalizing terms). The same remark applies to the rule (ans) in

Reasoning About Call-by-need by Means of Types

431

system A with respect to the one in [18]. Here is an example of typing derivation in system A.

It is worth noticing that system A is a conservative extension of system V: Lemma 4. Let t ∈ Ta . Then Φ V Γ t:σ iff Φ A Γ t:σ. A last remark of this section concerns relevance. Indeed, given Φ S Γ t:σ, not every free variable in t necessarily appears in the domain of Γ . More precisely, both systems enjoy the following property, that can be easily shown by induction on derivations. Lemma 5. Let S ∈ {V, A}. If Φ S Γ t:σ then dom(Γ ) ⊆ fv(t).

4

Characterization of name-normalizing Terms

In this section we adapt the (standard) characterization of head-normalizing terms of the λ-calculus to our call-by-name calculus, i.e. we show that a term t is name-normalizing iff t is typable in system V. The characterization is split in two parts. The first one shows that typable terms in system V are name-normalizing, a result which is based on a weighted subject reduction property. In contrast to similar results for idempotent intersection type systems, which are typically based on reducibility arguments, quantitative information of type derivations, obtained by means of non-idempotent types, provides a completely combinatorial proof. This is exactly the place where the quantitative approach by means of non-idempotent types makes the difference. The second part of the characterization shows that name-normalizing terms are typable in V, and is based on a subject expansion property. We start with the subject reduction property which uses the following lemma. Lemma 6 (Substitution). If Φt V x:{{σi }}i∈I ; Γ t:τ and (Φiu V Δi u:σi )i∈I then there exists a derivation Φt{x/u} s.t. Φt{x/u} V Γ +i∈I Δi t{x/u}:τ . Moreover, sz(Φt{x/u} ) = sz(Φt ) +i∈I sz(Φiu ) − |I|. Proof. By induction on Φt . Let Φt V x:{{σi }}i∈I ; Γ λy.u:a so that Γ = ∅ and I = ∅. Since Φt{x/u} V ∅ λy.u{x/u}:a, we conclude because ∅ = Γ +i∈∅ Δi . The other cases are similar to those of the λ-calculus (see for example [7]). ♦ Subject reduction can now be stated in the following form. Theorem 1 (Weighted Subject Reduction for name). Let Φ V Γ t:τ . If t →name t , then there exists Φ s.t. Φ V Γ t :τ . Moreover, sz(Φ) > sz(Φ ).

432

D. Kesner

Proof. By induction on Φ. The proof proceeds as that of the λ-calculus [7]. It is worth noticing that when t = (λx.u)u , the subterm λx.u is necessarily typed in Φ using rule (→ i), so that Lemma 6 can be applied to conclude. ♦ In order to show that name-normalizing terms are V-typable we use a subject expansion property, which needs the following lemma. Lemma 7 (Reverse Substitution). If Φt{x/u} V Γ t{x/u}:τ , then ∃Γ0 , ∃Φt , ∃I, ∃(Δi )i∈I , ∃(σi )i∈I , ∃(Φiu )i∈I such that Γ = Γ0 +i∈I Δi , and Φt V x : {{σi }}i∈I ; Γ0 t:τ and (Φiu V Δi u:σi )i∈I . Proof. By induction on Φt{x/u} . Suppose Φt{x/u} V ∅ λy.u:a so that Γ = ∅. There are two cases to consider: (1) t = λy.v and u = v{x/u}. Then Φt V ∅ λy.v:a, where Γ0 = ∅ and I = ∅ so that the property trivially holds. (2) t = x and u = λy.v. Then Φt V x : {{a}} x:a, Φu V λy.v:a, so that we let Γ0 = ∅, I = {1}, σ1 = a and Δ1 = ∅ and Γ = ∅ = Γ0 + Δ1 . All the other cases proceed as the one of the λ-calculus [7]. ♦ Theorem 2 (Subject Expansion for name). Let Φ V Γ t :τ . If t →name t , then there exists Φ s.t. Φ V Γ t:τ . Proof. By induction on Φ using Lemma 7.



We can now state the following full characterization result of this section: Theorem 3. Let t ∈ Ta . Then, Φ V Γ t:τ iff t ∈ WN (name). Proof. Let Φ V Γ t:τ . Then name-reduction necessarily terminates by Theorem 1 and thus t ∈ WN (name). Let t ∈ WN (name) so that t kname t , where t ∈ name-nf. We proceed by induction on k. If k = 0 (i.e. t = t ), then t is a name-nf. We have two cases. – If t = λy.v, then the property trivially holds with Γ = ∅ and τ = a. – If t = xt1 . . . tn (n ≥ 0), let Γ = {x:{{τ }}}, where τ = {{ }} → · · · → {{ }} → α (α is any base type and τ contains n occurrences of {{ }}). By the rule (ax) and n applications of (→ e) we obtain a derivation in V ending in Γ xt1 . . . tn :α. Otherwise, let t →name u kname t . By the i.h. we have a derivation in V ending in Γ u:τ . Thus by Theorem 2 the same holds form t. ♦

5

Characterization of need-normalizing Terms

This section gives a characterization of need-normalizing terms by means of A-typability, i.e. we show that a term t is need-normalizing iff t is typable in system A. At first sight, this result follows the same lines of Sect. 4, however, it is much trickier and requires the development of special tools dealing with need contexts. To the best of our knowledge, this is the first time that call-by-need normalization is characterized by means of typing technology.

Reasoning About Call-by-need by Means of Types

433

Similarly to the call-by-name case, we split the characterization proof in two parts and we use subject reduction and expansion properties to show them. More precisely, the first part of the characterization states that typable terms in system A are normalizing in call-by-need, a result based on a weighted subject reduction property, obtained by using quantitative information of the (non-idempotent) type derivations of system A. The second part of the characterization states that need-normalizing terms are typable in system A, a property based on the subject expansion property. In contrast to the call-by-name case, handling need-reduction is quite involved, particularly due to the use of contexts in the rewriting rules defining the relation. To deal with this difficulty, we introduce a technical tool which consists in extending system A (cf. Sect. 3) with typing rules for list contexts (Fig. 3). The type judgments have the form Γ  L  Δ, where Γ, Δ are type assignments and L is a list context. The left-hand side Γ of a judgment Γ  L  Δ is a type assignment for the (typed) free variables of L, while the right-hand side Δ is a type assignment for the term which is supposed to fill in the hole of L. These rules should then be considered modulo α-conversion. Notice the use of two different kinds of assignments in the second type rule: the assignments (Δi )i∈I are used to type the copies of u affecting the free occurrences of x in the list L, while (Δj )j∈J are used to type the copies of u affecting the free occurrences of x in the term which will fill the hole of L.

Fig. 3. Extending the intersection type system A to list contexts

The following lemma decomposes a type derivation of a term L[t] into one derivation for the context L and another one for the term t. Reciprocally, context and term derivations can be combined if their types coincide. The proof can be done by induction on L. Lemma 8. ΦL[t] A Λ L[t] : σ iff ∃Γ , ∃Π, ∃Δ, ∃ΦL , ∃Φt , such that Λ = Γ + Π and ΦL A Γ  L  Δ and Φt A Δ; Π t : σ. Moreover, sz(ΦL[t] ) = sz(ΦL ) + sz(Φt ) − 1. Combining different derivable typing judgments of the same list context by means of multiset union yields a derivable typing judgment. Moreover, their sizes can be related using the notion of height, defined on list contexts as follows: height() := 1 and height(L [y/v]) := height(L ) + 1.

434

D. Kesner

Lemma 9. If (ΦjL A Γj  L  Δj )j∈J , then there exists a derivation ΦL s.t. ΦL A +j∈J Γj  L  +j∈J Δj . Moreover, sz(ΦL ) = +j∈J sz(ΦjL ) − (height(L) · (|J| − 1)). Proof. Notice that the statement also holds in the case J = ∅. The proof is by induction on L (see the Appendix for details). ♦ To achieve the proof of the Weighted Subject Reduction property for the need-calculus, we first need to guarantee soundness of the (partial) substitution operation used in our framework, i.e. if N[[x]] and u are typable, then N[[u]] is typable too. Moreover, |K| elements of the multiset type {{σi }}i∈I associated to x in the type derivation of N[[x]] are consumed, for some K ⊆ I. Lemma 10 (Partial Substitution). If ΦN[[x]] A x:{{σi }}i∈I ; Γ N[[x]]:τ and (Φiu A Δi u:σi )i∈I then ∃ΦN[[u]] s.t. ΦN[[u]] A x:{{σi }}i∈IK ; Γ +k∈K Δk N[[u]]:τ , for some K ⊆ I where sz(ΦN[[u]] ) = sz(ΦN[[x]] ) +k∈K sz(Φku ) − |K|. Proof. By induction on typing derivations (a similar proof appears in [17]).



To complete the subject reduction argument we still need to guarantee that every needed variable of a typed term is necessarily typed, a property which is specified by means of the following lemma. Lemma 11. If Φ A Γ N[[x]]:τ , then ∃Γ  , ∃I = ∅, and ∃(σi )i∈I such that Γ = x : {{σi }}i∈I ; Γ  . Proof. By induction on N. The base case N =  is straightforward. We only show here the interesting inductive case N[[x]] = N1 [[y]][y/N2 [[x]]], for which Φ has necessarily the following form, where Γ = Π +j∈J Δj .

By the i.h. on the left derivation we have J = ∅. By the i.h. on the right derivations Δj = x : {{τji }}i∈Ij ; Δj and Ij = ∅ for all j ∈ J. Moreover, Π = x : {{τk }}k∈K ; Π  , where K could be empty. Then Π +j∈J Δj = (x : {{τk }}k∈K ; Π  ) +j∈J (x : {{τji }}i∈Ij ; Δj ) = x : {{τk }}k∈K +j∈J x : {{τji }}i∈Ij ; (Π  +j∈J Δj ) The property then holds for {{σi }}i∈I = {{τk }}k∈K +j∈J {{τji }}i∈Ij and Γ  = ♦ Π +j∈J Δj . We have I = ∅ since J = ∅ and Ij = ∅ for all j ∈ J. 

Subject reduction can now be stated in a combinatorial way, by using the sz( ) measure on derivations defined in Sect. 3. Theorem 4 (Weighted Subject Reduction for need). Let Φ A Γ t:τ . If t →need t , then there exists Φ s.t. Φ A Γ t :τ . Moreover, sz(Φ) > sz(Φ ). Proof. By induction on t →need t , using Lemmas 8, 9, 10 and 11. We refer the interested reader to the Appendix for full details of the proof. ♦

Reasoning About Call-by-need by Means of Types

435

To show that need-normalizing terms are A-typable we use a subject expansion property, which needs the following lemma. Lemma 12 (Reverse Partial Substitution). Let N[[x]], s be terms s.t. x ∈ / fv(s) and ΦN[[s]] A Γ N[[s]]:τ . Then ∃Γ0 , ∃ΦN[[x]] , ∃I, ∃(Δi )i∈I , ∃(σi )i∈I , ∃(Φis )i∈I s.t. Γ = Γ0 +i∈I Δi and ΦN[[x]] A x : {{σi }}i∈I + Γ0 N[[x]]:τ and (Φis A Δi s:σi )i∈I . Proof. By induction on the structure of N[[s]]. – If N = , then N[[s]] = s and the result holds, for Γ0 = ∅, |I| = 1 and σ1 = τ . – If N = λy.M then the property is straightforward by the i.h. (since y ∈ / fv(s) by α-conversion). – If N = Mr then N[[s]] = M[[s]]r and by construction Γ = Π +j∈J Γj and ΦM[[s]]  Π M[[s]]:{{ρj }}j∈J → τ and (Φjr  Γj r:ρj )j∈J . By the i.h. Π = Π0 +i∈I Δi where x : {{σi }}i∈I + Π0 M[[x]]:{{ρj }}j∈J → τ and (Δi s:σi )i∈I . Then, by the rule (→ e), x : {{σi }}i∈I + Π0 +j∈J Γj M[[x]]r:τ . The result then holds for Γ0 := Π0 +j∈J Γj . – If N = rM then N[[s]] = rM[[s]] and by construction Γ = Π +j∈J Γj and Φr  Π r:{{ρj }}j∈J → τ and (ΦjM[[s]]  Γj M[[s]]:ρj )j∈J . By the i.h. for each j ∈ J,

Γj = Γ0j +i∈Ij Γij where x : {{σi }}i∈Ij + Γ0j M[[x]]:ρj and (Γij s:σi )i∈Ij . Let I := ∪j∈J Ij . Then, by the rule (→ e), Π +j∈J (x : {{σi }}i∈Ij + Γ0j ) rM[[x]]:τ . Note that Π +j∈J (x : {{σi }}i∈Ij + Γ0j ) = x : {{σi }}i∈I + Π +j∈J Γ0j . The result then holds for Γ0 := Π +j∈J Γ0j . – All the remaining cases are similar to the previous ones. ♦

Theorem 5 (Subject Expansion for need). Let Φ A Γ t :τ . If t →need t , then there exists Φ s.t. Φ A Γ t:τ . Proof. By induction on Φ using Lemmas 8, 9 and 12.



To state the full characterization result of this section, we still need to relate typing of Te -terms in system A with typing of Ta -terms in system V. Lemma 13. Let t ∈ Te . Then Φ V Γ t↓ :σ implies Φ A Γ t:σ. Proof. By induction on t. – If t = x, then t↓ = x. The derivation Φ has the form x : {{σ}} x:σ so that Φ has also the form x : {{σ}} x:σ. – If t = λx.u, then t↓ = λx.u↓ . We have Γ = Δ \\ x and σ = Δ(x) → τ , where Δ \\ x λx.u↓ :Δ(x) → τ is necessarily derivable from a derivation Φu V Δ u↓ :τ . The i.h. gives Φu A Δ u:τ so that we construct Φ A Δ \\ x λx.u:Δ(x) → τ which concludes the proof. – If t = t1 t2 , then the proof proceeds by induction as in the previous case. – If t = t1 [x/t2 ], then t↓ = t1 ↓ {x/t2 ↓ }. By Lemma 7 we know that Γ = Γ0 +i∈I Δi and Φt1 ↓ V x : {{σi }}i∈I ; Γ0 t1 ↓ :τ and (Φit2 ↓ V Δi t2 ↓ :σi )i∈I . The i.h. then gives Φt1 A x : {{σi }}i∈I ; Γ0 t1 :τ and (Φi t2 A Δi t2 :σi )i∈I . We then ♦ conclude Φt A Γ t1 [x/t2 ]:τ by using rule (cut).

436

D. Kesner

We can now state the full characterization result, which is one of the main results of this paper. The implication “typable terms in system A are neednormalizing” is obtained through the weighted subject reduction, which provides a completely combinatorial argument of this property, as in the call-by-name case. This is exactly the advantage provided by the quantitative approach based on non-idempotent types, which makes the difference with a qualitative system based on idempotent intersection types. Theorem 6. Let t ∈ Te . Then, Φ A Γ t:τ iff t ∈ WN (need). Proof. Let ΦΓ t:τ . Then need-reduction necessarily terminates by Theorem 4 so that t ∈ WN (need). Let t ∈ WN (need), so that t kneed t , where t ∈ need-nf. We proceed by induction on k. If k = 0 (i.e. t = t ), then t is in need-nf. By Lemma 3 t↓ is in name-nf. Then Φ V Γ t↓ :σ by Theorem 3 and Φ A Γ t:σ by Lemma 13. Otherwise, let t →need u →kneed t . By the i.h. we have Φ A Γ u:τ , thus by ♦ Theorem 5 we obtain Φ A Γ t:τ .

6

Soundness and Completeness

This section uses the two characterization results developed in Sects. 4 and 5 to prove soundness and completeness of call-by-need w.r.t call-by-name. More precisely, a call-by-name interpreter stops in a value if and only if the call-by-need interpreter stops in an answer. This implies that call-by-need and call-by-name are observationally equivalent, so that in particular call-by-need turns out to be a correct implementation of call-by-name. Lemma 14. Let t ∈ Ta . Then t ∈ WN (need) ⇔ t ∈ WN (name). Proof. Let t ∈ WN (need). Then Φ A Γ t:τ holds by Theorem 6, and Φ V Γ t:τ holds by Lemma 4, thus we get t ∈ WN (name) by Theorem 3. Conversely, let t ∈ WN (name). Then Φ V Γ t:τ holds by Theorem 3, so that Φ A Γ t:τ holds by Lemma 4, and thus we get t ∈ WN (need) by Theorem 6. ♦ Given a reduction relation R on a term language T , and an associated notion of context for the term language T , we define t to be observationally equivalent to u, written t ∼ =R u, if and only if C[t] ∈ WN (R) ⇔ C[s] ∈ WN (R) for every context C. Since we work with two different term languages Ta and Te , we first need to introduce their associated notions of contexts C and C , which represent, respectively, contexts without and with explicit substitutions. C ::=  | λx.C | C t | t C C ::=  | λx.C | C t | t C | C [x/t] | t[x/C ] We can thus conclude with the last main result of the paper.

Reasoning About Call-by-need by Means of Types

437

Corollary 1. For all terms t and u in Ta , t ∼ =name u iff t ∼ =need u. Proof. First of all let us consider the relation =B , which is the least equivalence relation generated by the axiom (λx.t1 )t2 = t1 [x/t2 ]. We remark that when t is typable in A, and t =B t , then t is typable in A. This means that we can indistinctly quantify over the two sort of contexts C and C previously defined: indeed, from the set of all the contexts C we can construct the set of all the contexts C , and vice-versa. The proof of the corollary then proceeds as follows. Take t, u ∈ Ta . Then t ∼ =name u iff (definition) C[t] ∈ WN (name) ⇔ C[u] ∈ WN (name) for every context C iff (Theorem 3) C[t] is typable in V ⇔ C[u] is typable in V for every context C iff (Lemma 4) C[t] is typable in A ⇔ C[u] is typable in A for every context C iff (Remark) C [t] is typable in A ⇔ C [u] is typable in A for every context C iff (Theorem 6) C [t] ∈ WN (need) ⇔ C [u] ∈ WN (need) for every context C iff (definition) ♦ t∼ =need u.

7

Conclusion

This paper gives the first full characterization of call-by-need normalization by means of intersection types. This result, together with the full characterization of call-by-name normalization, provides a new completeness proof for call-byneed, which is only based on logical arguments and does not make use of any complicated notion of rewriting. The use of non-idempotent types allows us to obtain the result in a combinatorial way, by providing quantitative information about reduction sequences, and without resorting to any reducibility argument. The paper only considers a core language for call-by-need, but it would be interesting to consider other constructors of lazy languages, e.g. data constructors, case-expressions and Haskell’s seq operators. Moreover, the results in the paper could be extended to the cyclic call-by-need letrec calculus as well as to full (strong) need normal forms (ongoing work). Last but not least, we would also like to use the ideas in [16] in order to relate call-by-need and needed reduction by means of intersection types.

A

Appendix

Lemma 9. If (ΦjL A Γj  L  Δj )j∈J , then ΦL A +j∈J Γj  L  +j∈J Δj . Moreover, sz(ΦL ) = +j∈J sz(ΦjL ) − (height(L) · (|J| − 1)). Proof. The proof is by induction on L. The case L =  is straightforward since Γj = Δj = ∅ so that +j∈J Γj = +j∈J Δj = ∅. We have sz(Φ ) = 1 = |J| − |J| + 1 = |J| − [1 · (|J| − 1)] = +j∈J sz(Φj ) − (height() · (|J| − 1)). Let L = L [y/u]. Then for each j ∈ J, the derivation ΦjL has the following form, where Γj = Πj +i∈Kj Hj Λjk and Δj = Δj ; y : {{σh }}h∈Hj .

438

D. Kesner

By the i.h. we have ΦL +j∈J (y : {{σk }}k∈Kj ; Πj )  L  +j∈J Δj , but +j∈J (y : {{σk }}k∈Kj ; Πj ) = y : {{σk }}j∈J,k∈Kj ; +j∈J Πj . We thus construct the derivation ΦL :

Λjk

We conclude with the first statement since +j∈J Γj = +j∈J Πj +j∈J,i∈Kj Hj and +j∈J Δj = +j∈J Δj , y : {{σh }}j∈J,h∈Hj . Moreover: sz(ΦL ) = sz(ΦL ) +j∈J,i∈Kj Hj sz(ΦΨj,i ) + 1 =i.h. u +j∈J sz(ΦL ) − (height(L ) · (|J| − 1)) +j∈J,i∈Kj Hj sz(ΦΨj,i ) + 1 =i.h. u +j∈J sz(ΦL ) +j∈J,i∈Kj Hj sz(ΦΨj,i ) + |J| − [(height(L ) + 1) · (|J| − 1)] = u +j∈J sz(ΦL ) − (height(L) · (|J| − 1)) ♦

Theorem 4 (Weighted Subject Reduction for need). Let Φ A Γ t:τ . If t →need t , then there exists Φ s.t. Φ A Γ t :τ . Moreover, sz(Φ) > sz(Φ ). Proof. By induction on t →need t . – If t = L[λx.u]s →dB L[u[x/s]] = t , then one shows Φ  Γ t :τ and sz(Φ) > sz(Φ ) by induction on L. See the first case of Lemma 2 in [17] (pp. 19). – If t = N[[x]][x/L[u]] →lsv L[N[[u]][x/u]] = t , then the derivation Φ has the following form, where Γ = Γ0 +i∈I Δi .

By Lemma 8, for all i ∈ I, there exist Π1i , Π2i , Π3i such that ΦiL  Π1i  L  Π2i , Φiu  Π2i ; Π3i u : σi and Δi = Π1i + Π3i . Then, from the derivations ΦN[[x]] and (Φiu )i∈I we get, by Lemma 10, a derivation ΦN[[u]]  x : {{σi }}i∈I\K ; Γ0 +k∈K (Π2k ; Π3k ) N[[u]] : σ for some K ⊆ I. Therefore, we can construct the following derivation ΦN[[u]][x/u] .

The last sequent can be written Γ0 + (+i∈I Π2i ; +i∈I Π3i ) N[[u]][x/u] : σ. Now, from ΦN[[x]] and Lemma 11 we know that I = ∅. We can thus apply Lemma 9 to (ΦiL )i∈I and we get ΦL  +i∈I Π1i  L  +i∈I Π2i . We can thus apply Lemma 8 to ΦL and ΦN[[u]][x/u] , obtaining Φ  Γ0 +i∈I Π11 +i∈I Π3i L[N[[u]][x/u]] : σ.

Reasoning About Call-by-need by Means of Types

439

We can then conclude with the first statement since Γ0 +i∈I Π11 +i∈I Π3i = Γ0 +i∈I Δi = Γ as required. Moreover, for the second one, we have two equalities: =L.8 sz(Φ) = sz(ΦN[[x]] ) +i∈I sz(ΦiL[u] ) + 1 sz(ΦN[[x]] ) +i∈I (sz(ΦiL ) + sz(Φiu ) − 1) + 1 = sz(ΦN[[x]] ) +i∈I sz(ΦiL ) +i∈I sz(Φiu ) − (|I| − 1) = Z − (|I| − 1) and sz(Φ ) =L.8 sz(ΦL ) + sz(ΦN[[u]][x/u] ) − 1 = = sz(ΦL ) + sz(ΦN[[u]] ) +i∈I\K sz(Φiu ) + 1 − 1 =L.10 sz(ΦL ) + sz(ΦN[[u]] ) +i∈I\K sz(Φiu ) sz(ΦL ) + sz(ΦN[[x]] ) +k∈K sz(Φku ) − |K| +i∈I\K sz(Φiu ) =L.9 +i∈I sz(ΦiL ) − [height(L) · (|I| − 1)] + sz(ΦN[[x]] ) +i∈I sz(Φiu ) − |K| = sz(ΦN[[x]] ) +i∈I sz(ΦiL ) +i∈I sz(Φiu ) − [height(L) · (|I| − 1)] − |K| = Z − [height(L) · (|I| − 1)] − |K| To conclude, we know by Lemma 10 that K = ∅. Therefore, |I| − 1 ≤ height(L) · (|I| − 1) so that Z − (|I| − 1) ≥ Z − [height(L) · (|I| − 1)] > Z − [height(L) · (|I| − 1)] − |K|. We thus conclude sz(Φ) > sz(Φ ) as required. – If t = N1 [[x]][x/N2 [u]] →need N1 [[x]][x/N2 [u ]] = t comes from N[u] →need N[u ], then Φ has the following form, where Γ = Γ0 +i∈I Δi .

By the i.h. both ΠNi2 [u]  Δi N2 [u ]:σi and sz(ΦiN2 [u] ) > sz(ΠNi2 [u] ) hold for all i ∈ I. The first statement gives Φ  Γ0 +i∈I Δi N1 [[x]][x/N2 [u ]]:τ . The second one gives sz(Φ) > sz(Φ ) since I = ∅ holds by Lemma 11. – The cases t = N[u]v →need N[u ]v = t and t = N[u][x/v] →need N[u ][x/v] = t coming from N[u] →need N[u ] are straightforward inductive cases. ♦

References 1. Accattoli, B., Barenbaum, P., Mazza, D.: Distilling abstract machines. In: ICFP. ACM Press (2014) 2. Ariola, Z.M., Felleisen, M.: The call-by-need lambda calculus. J. Funct. Program. 7(3), 265–301 (1997) 3. Ariola, Z.M., Felleisen, M., Maraist, J., Odersky, M., Wadler, P.: The call-by-need lambda calculus. In: POPL, pp. 233–246. ACM Press (1995) 4. Barendregt, H., Coppo, M., Dezani-Ciancaglini, M.: A filter lambda model and the completeness of type assignment. Bull. Symbolic Logic 48, 931–940 (1983) 5. Bernadet, A., Lengrand, S.: Non-idempotent intersection types and strong normalisation. Logical Methods Comput. Sci. 9(4:3), 1–46 (2013) 6. Boudol, G., Curien, P.-L., Lavatelli, C.: A semantics for lambda calculi with resources. Math. Struct. Comput. Sci. 9(4), 437–482 (1999)

440

D. Kesner

7. Bucciarelli, A., Kesner, D., Ronchi Della Rocca, S.: The inhabitation problem for non-idempotent intersection types. In: Diaz, J., Lanese, I., Sangiorgi, D. (eds.) TCS 2014. LNCS, vol. 8705, pp. 341–354. Springer, Heidelberg (2014) 8. Cardone, F., Coppo, M.: Two extension of Curry’s type inference system. In: Odifreddi, P. (ed.) Logic and Computer Science. APIC Series, vol. 31, pp. 19–75. Academic Press, New York (1990) 9. Chang, S., Felleisen, M.: The call-by-need lambda calculus, revisited. In: Seidl, H. (ed.) ESOP 2012. LNCS, vol. 7211, pp. 128–147. Springer, Heidelberg (2012) 10. Coppo, M., Dezani-Ciancaglini, M., Venneri, B.: Functional characters of solvable terms. Math. Logic Q. 27(2–6), 45–58 (1981) 11. Coppo, M., Dezani-Ciancaglini, M.: A new type-assignment for lambda terms. Archiv f¨ ur Math. Logik Grundlagenforschung 19, 139–156 (1978) 12. Coppo, M., Dezani-Ciancaglini, M.: An extension of the basic functionality theory for the λ-calculus. Notre Dame J. Formal Logic 21, 685–693 (1980) 13. Danvy, O., Zerny, I.: A synthetic operational account of call-by-need evaluation. In: PPDP, pp. 97–108. ACM Press (2013) 14. De Benedetti, E., Ronchi Della Rocca, S.: Bounding normalization time through intersection types. In: ITRS 2012, EPTCS, pp. 48–57 (2013) 15. de Carvalho, D.: S´emantiques de la logique lin´eaire et temps de calcul. Th`ese de doctorat, Universit´e Aix-Marseille II (2007) 16. Gardner, P.: Discovering needed reductions using type theory. In: Hagiya, M., Mitchell, J.C. (eds.) TACS 1994. LNCS, vol. 789, pp. 555–574. Springer, Heidelberg (1994) 17. Kesner, D., Ventura, D.: Quantitative types for intuitionistic calculi. Technical report (2014). https://hal.archives-ouvertes.fr/hal-00980868 18. Kesner, D., Ventura, D.: Quantitative types for the linear substitution calculus. In: Diaz, J., Lanese, I., Sangiorgi, D. (eds.) TCS 2014. LNCS, vol. 8705, pp. 296–310. Springer, Heidelberg (2014) 19. Kfoury, A.J.: A linearization of the lambda-calculus and consequences. Technical report, Boston Universsity (1996) 20. Kfoury, A.J.: A linearization of the lambda-calculus and consequences. J. Logic Comput. 10(3), 411–436 (2000) 21. Kfoury, A.J., Wells, J.B.: Principality and type inference for intersection types using expansion variables. Theor. Comput. Sci. 311(1–3), 1–70 (2004) 22. Krivine, J.-L.: Lambda-Calculus, Types and Models. Ellis Horwood, Hemel Hempstead (1993) 23. Mairson, H., Neergaard, P.M.: Types, potency, idempotency: why nonlinearity and amnesia make a type system work. In: ICFP, pp. 138–149. ACM (2004) 24. Maraist, J., Odersky, M., Turner, D.N., Wadler, P.: Call-by-name, call-by-value, call-by-need and the linear lambda calculus. Theoret. Comput. Sci. 228(1–2), 175– 210 (1999) 25. Maraist, J., Odersky, M., Wadler, P.: The call-by-need lambda calculus. J. Funct. Program. 8(3), 275–317 (1998) 26. Pagani, M., Ronchi Della Rocca, S.: Linearity, non-determinism and solvability. Fundam. Informaticae 103, 358–373 (2010) 27. Pagani, M., della Rocca, S.R.: Solvability in resource lambda-calculus. In: Ong, L. (ed.) FOSSACS 2010. LNCS, vol. 6014, pp. 358–373. Springer, Heidelberg (2010)

Reasoning About Call-by-need by Means of Types

441

28. Plotkin, G.D.: Call-by-Name, call-by-value and the lambda-calculus. Theor. Comput. Sci. 1(2), 125–159 (1975) 29. van Bakel, S.: Complete restrictions of the intersection type discipline. Theor. Comput. Sci. 102(1), 135–163 (1992) 30. Wadsworth, C.: Semantics and Pragmatics of the Lambda Calculus. PH.D. thesis, Oxford University (1971)

Algorithms for Infinite Systems

Coverability Trees for Petri Nets with Unordered Data Piotr Hofman1 , Slawomir Lasota2 , Ranko Lazi´c3 , J´erˆ ome Leroux4 , 1(B) , and Patrick Totzke3 Sylvain Schmitz 1

3

LSV, ENS Cachan & CNRS, Universit´e Paris-Saclay, Cachan, France [email protected] 2 University of Warsaw, Warsaw, Poland DIMAP, Department of Computer Science, University of Warwick, Coventry, UK 4 LaBRI, CNRS, Paris, France Abstract. We study an extension of classical Petri nets where tokens carry values from a countable data domain, that can be tested for equality upon firing transitions. These Unordered Data Petri Nets (UDPN) are well-structured and therefore allow generic decision procedures for several verification problems including coverability and boundedness. We show how to construct a finite representation of the coverability set in terms of its ideal decomposition. This not only provides an alternative method to decide coverability and boundedness, but is also an important step towards deciding the reachability problem. This also allows to answer more precise questions about the reachability set, for instance whether there is a bound on the number of tokens on a given place (place boundedness), or if such a bound exists for the number of different data values carried by tokens (place width boundedness). We provide matching Hyper-Ackermann bounds on the size of coverability trees and on the running time of the induced decision procedures.

1

Introduction

Unordered data Petri nets (UDPN [15]) extend Petri nets by decorating tokens with data values taken from some countable data domain D. These values act as pure names: they can only be compared for equality or non-equality upon firing transitions. Such systems can model for instance distributed protocols where process identities need to be taken into account [21]. UDPNs also coincide with the natural generalisation of Petri nets in the framework of sets with atoms [3]. In spite of their high expressiveness, UDPNs fit in the large family of Petri net P. Hofman—Partially funded by the Polish National Science Centre grant 2013/09/B/ST6/01575 and by Labex Digicosme, Universit´e Paris-Saclay, project VERICONISS. S. Lasota—Partially funded by the Polish National Science Centre grant 2012/07/B/ST6/01497. R. Lazi´c and P. Totzke—Partially supported by the EPSRC grant EP/M011801/1. S. Schmitz—Partially supported by the ANR grant ANR-14-CE28-0005 prodaq and by the Leverhulme Trust Visiting Professorship VP1-2014-041. c Springer-Verlag Berlin Heidelberg 2016  B. Jacobs and C. L¨ oding (Eds.): FOSSACS 2016, LNCS 9634, pp. 445–461, 2016. DOI: 10.1007/978-3-662-49630-5 26

446

P. Hofman et al.

Fig. 1. A short taxonomy of some well-structured extensions of Petri nets. Complexities in violet refer to the coverability and termination problems, and can be taken as proxies for expressiveness; the exact complexities of coverability and termination in ν-Petri nets and UDPNs are unknown at the moment. Place boundedness is decidable below the yellow line and undecidable above. As indicated by the dashed arrows, freshness can be enforced using a dense linear order or whole-place operations (Colour figure online).

extensions among the well-structured ones [1, 7]. As such, they still enjoy decision procedures for several verification problems, prominently safety (through the coverability problem) and termination. Unordered data Petri nets have an interesting position in the taxonomy of well-structured Petri net extensions (see Fig. 1). Indeed, all their extensions forgo the decidability of the reachability problem (whether a target configuration is reachable) and of the place boundedness problem (whether the number of tokens in a given place can be bounded along all runs): this is the case of ν-Petri nets [21] that allow to create fresh data values, of ordered data Petri nets [15] that posit a dense linear ordering on D, and of unordered data nets [15] that allow to perform ‘whole-place’ operations, which move and/or duplicate all the tokens from a place to another. By contrast, it is currently open whether reachability is decidable in UDPNs, and a consequence of our results in this paper is that place boundedness is decidable—which is a significant first step if we wish to adapt to UDPNs some of the known algorithms for reachability in Petri nets [13, 18]. Contributions. In this paper, we show how to construct finite coverability trees for UDPNs, adapting the existing construction of Karp and Miller [12] for Petri nets. Such trees are constructed forward from an initial configuration like reachability trees, but approximate the latter by accelerating sequences of transitions and explicitly manipulating limits of reachable configurations as downwardsclosed sets (see Sect. 4.1). We rely for all this on a general theory for representing downwards-closed sets as finite unions of ideals developed by Finkel and Goubault-Larrecq [9] for this exact purpose (see Sect. 3).

Coverability Trees for Petri Nets with Unordered Data

447

Coverability trees contain a wealth of information about the system at hand, and allow to answer various coverability and boundedness questions—allowing us to derive a new result: the place boundedness problem is decidable for UDPNs, and so are its variants, like place width- and place depth boundedness (see Sect. 2). We also establish in Sect. 5 matching ‘hyper-Ackermannian’ lower and upper bounds on the size of UDPNs coverability trees. This yields Fωω upper bounds for the already mentioned decidable problems in UDPNs, in terms of the fast-growing complexity classes (Fα )α from [23]. These complexity results rely largely on the work of Rosa-Velardo [20] on the complexity of coverability in unordered data nets. Due to space constraints, most proof details are omitted but can be found online in the full version of the paper available from https:// hal.inria.fr/hal-01252674. Related Work. A coverability tree construction has already been undertaken by Rosa-Velardo, Martos-Salgado, and de Frutos-Escrig [22] in the case of ν-Petri nets, and inescapably there are many similarities between their work and ours. Our construction does however not merely remove freshness constraints from theirs: (1) we start a new and rely on a strong invariant on the form of ideals in UDPN coverability trees, which leads to significant simplifications but would be markedly difficult to extract from Rosa-Velardo et al.’s [22] construction, and (2) coverability trees are not necessarily finite for ν-Petri nets (place boundedness is indeed undecidable [21]), which means that our termination argument and complexity bounds are entirely new considerations. Like Rosa-Velardo et al. [22] we rely on the work of Finkel and GoubaultLarrecq [9] on forward analysis of well-structured systems. Finkel and GoubaultLarrecq provide in particular an abstract generic procedure, but without guarantee of termination in general, and their framework needs to be instantiated for each specific class of systems.

2

Model

Our presentation of unordered data Petri nets differs from the original one [15] on two counts: we work with an equivalent formalism with more of a vector addition system [12] flavour, and because we need to work with ideals we define the syntax and the semantics on extended configurations, which allow for infinitely many different data values and infinite counts. Let Z and N denote the sets of integers and non-negative integers respectively, def def and complete them as Zω = Z  {ω} and Nω = N  {ω} with a new top element ω with ω > z and z + ω = ω + z = ω for all z in Z. Given a dimension k in N, we denote the projection into the ith component of a vector v ∈ Zkω by v[i] and define the product ordering and sum over Zkω componentwise: u ≤ v if def

u[i] ≤ v[i] for all 1 ≤ i ≤ k, and (u + v)[i] = u[i] + v[i] for all 1 ≤ i ≤ k. We write 0 for the vector with 0 on all components. Data Vectors. Fix some countable domain D of data values and a dimension k in N. A data vector is a function f : D → Zkω . Data vectors can be partially ordered

448

P. Hofman et al. def

and summed pointwise: f ≤ g if f (d) ≤ g(d) for all d in D, and (f + g)(d) = f (d) + g(d) for all d in D. As usual we write f < g if f ≤ g and f (d) < g(d) for some d in D. For a subset K ⊆ {1, . . . , k} we write f K for the projection of f def

into components in K: for all d in D, f K (d) = (f (d))K . The support of a data def

vector f is Supp 0 (f ) = {d ∈ D | f (d) = 0}; f is finitely supported if Supp 0 (f ) is finite. We say that a data vector f is non-negative if f (d) belongs to Nkω for all d in D. It is finite if f (d) belongs to Zk for all d and it is finitely supported. We call bijections σ : D → D (data) permutations and write f σ for the composition of a permutation σ and a data vector f . If two data vectors f, g satisfy f = gσ for some permutation σ, we say f and g are equal up to permutation and write f ≡ g. In the sequel we will consider sets of data vectors that are finite up to permutation: a set X of data vectors is finite up to permutation if there is a finite subset X  of X such that every f ∈ X is equal up to permutation to some f  ∈ X  . Note that if X is closed under permutations then X  can be used as its finite presentation; any such finite set X  we call representative of X. Definition 2.1 (UDPN). An unordered data Petri net (UDPN) is a finite set def

T of finite data vectors. A transition is a data vector t = f σ, where f ∈ T t and σ is a data permutation. There is a step f − → g between non-negative data vectors f, g if g = f + t for some transition t. Note that this enforces that f (d) + t(d) ≥ 0 for all d in D since g is non-negative. We simply write f − → g if t ∗ → g for some transition t and let − → denote the transitive and reflexive closure f− of − →. A configuration is a finite non-negative data vector f , i.e. f (d) belongs to Nk for all d in D, and f (d) = 0 for almost all d in D. We write Confs for the set of configurations and note that Confs is closed under UDPN steps. The reachability set from a given vector f is defined as def



→ g}. Reach(f ) = {g ∈ Confs | f −

(1)

Observe that UDPNs over any domain with cardinality |D| = 1 are classical vector addition systems [12]. Notice also that, the set of transitions in an UDPN is finite up to permutations of D and that the step relation is closed under permutations: For every non-negative data vector f , transition t and permutation σ we have that t



→ g implies f σ −→ gσ. f− def

(2)

def

Example 2.2. For the domain D = N and k = 2, consider a 2-dimensional UDPN T = {t1 , t2 }, with vectors t1 , t2 defined as t1 : 0 → (1, 0) and t1 : n → (0, 0) for all n > 0, and t2 : 0 → (−1, −1), t2 : 1 → (1, 1) and t2 : n → (0, 0) for all n > 1. The configuration f0 with f0 : 0 → (1, 1) and f0 : n → (0, 0) for n > 0, def

has infinitely many t1 -successors. Namely, gi = f0 + t1 σi for every permutation

Coverability Trees for Petri Nets with Unordered Data

449

Fig. 2. A place/transition representation of the UDPN of Example 2.2 in the style of [11, 21]. Different data values are depicted though differently coloured tokens in places (the circles), and through differently named variables in transitions (the boxes and arrows).

σi that swaps 0 and i ∈ D. However, there are only two such successors up to permutation because gi ≡ gj for all i, j > 0. The reachability set of f0 is ⎧ ⎫ ⎨ ∃d1 , d2 , . . . , dm ∈ D ∃n1 , n2 , . . . , nm ∈ N ⎬ Reach(f0 ) = g g(d1 ) = (n1 , 1) and ∀1 < i ≤ m, g(di ) = (ni , 0), . (3) ⎩ ⎭ ∀d ∈ D \ {d1 , d2 , . . . , dm } g(d) = (0, 0) In the sequel, we present UDPNs only up to permutation in matrix form def  def by juxtaposing the vectors from their finite supports: we write t1 = 10 , t2 =  −1 1    def 1 2 1 1 and f0 = 1 . The t1 -successors of f0 are g0 = 1 and gi = 1 +  −1 1 = 1 1 for i = 0. The latter is depicted in Fig. 2 in the style of coloured 0 10 place/transition

 n0 n1 ··· nm nets. The reachability set of f0 can be written as Reach(f0 ) = | m ≥ 0, n0 , . . . , nm ≥ 1 . 1 0 ··· 0 Embeddings. We say that a data vector f embeds into a data vector g and write f g (resp. f  g) if there exists an injection π: D → D such that f ≤ gπ (resp. f < gπ). The injection π itself is called an embedding (of f into g) and a permutation embedding in case it is bijective. Given a set of configurations C, its downward-closure ↓C is {f ∈ Confs | ∃g ∈ C . f g}, and as usual a set C is downwards-closed if ↓C = C. Finitely supported data vectors are isomorphic to finite multisets of vectors in Zkω when working up to data permutation. Moreover, on permutation classes of finitely supported data vectors, the embedding ordering coincides with the usual embedding ordering over finite multisets of vectors; in consequence, UDPN configurations are well-quasi-ordered by the embedding ordering. Thus an UDPN defines a quasi-ordered transition system (Confs, − →, ), which satisfies a (strong) compatibility condition as shown in the following lemma. Together with the fact that (Confs, ) is a wqo, this entails that it is a well-structured transition system in the sense of [1, 7]. Lemma 2.3. (Strong Strict Compatibility). Let f, f  , g be configurations. → g, then there exists a configuration g  with If f f  (resp. f  f  ) and f −     → g and g g (resp. g  g ). f − t

→ g, and a permutation Proof. Consider a finite data vector t such that f − π of D such that f ≤ f  π (recall that, when working with finitely supported data vectors, embeddings can be assumed to be permutations). We claim that

450 def

P. Hofman et al. tπ −1

g  = f  + tπ −1 satisfies f  −−−→ g  and g ≤ g  π. Indeed, for all d in D, noting def e = π(d), g(d) = f (d) + t(d) ≤ f  (π(d)) + t(d) = f  (e) + t(π −1 (e)) = g  (e) = g  (π(d)) . Furthermore, assuming f < f  π, for at least one d in D the above inequality is strict.   Decision Problems. For the purpose of verification, we are interested in standard ∗ → g hold for given decision problems for UDPN, including reachability (does f − configurations f and g?), coverability (given configurations f, g, does there exists ∗ g   g s.t. f − → g  ?), and boundedness (is Reach(f ) finite up to permutation?). While the decidability of reachability remains open, well-structuredness of UDPN (and some basic effectiveness assumptions) implies that the coverability and boundedness problems are decidable using the generic algorithms from [1,7]. In fact, decidability holds more generally for ordered data Petri nets [15, Theorm 4.1]. For the coverability problem, Rosa-Velardo [20, Theorem 1] proved an HyperAckermann upper bound (Fωω in the hierarchy from [23]; see Sect. 5), while Lazi´c et al. [15, Theorem 5.2] proved a Tower lower bound (F3 in the same hierarchy). The complexity of the boundedness problem has not been studied before. Furthermore, the following, more precise variant of boundedness called place boundedness was not known to be decidable: Input: An UDPN, a configuration f , and a set of coordinates K ⊆ {1, . . . , k}. Question: Is {gK | g ∈ Reach(f )} finite up to permutation? In the presence of an infinite data domain D, place boundedness can be further refined: even if infinitely many configurations are reachable, the system can still be bounded in the sense that there exists a bound on the number of different data values in reachable configurations; Rosa-Velardo and de Frutos-Escrig [21] call this bounded width. Similarly, there may exist some bound on the multiplicities with which any data value occurs, while the number of different data values is unbounded; Rosa-Velardo and de Frutos-Escrig [21] call this bounded depth. We formalise the resulting decision problems in our notation as follows. The place width boundedness problem is given as: Input: An UDPN, a configuration f , and a set of coordinates K ⊆ {1, . . . , k}. Question: Is {|Supp 0 (gK )| | g ∈ Reach(f )} finite? The place depth boundedness problem is the following: Input: An UDPN, a configuration f , and a set of coordinates K ⊆ {1, . . . , k}. Question: Is {gK (d) | g ∈ Reach(f ), d ∈ D} finite? If the answer to the depth (width) boundedness problem is positive we call those components i ∈ K depth (width) bounded.

Coverability Trees for Petri Nets with Unordered Data

451

 Example 2.4. From the initial configuration f0 =  11 , the UDPN from Example 2.2 can reach any configuration of the form n1 in n many t1 -steps, all on the same data value. Similarly, any configuration of the form  1 1exercised ··· 1 1 0 ··· 0 with a support of size n + 1 can be reached after a sequence of n transitions t1 , each exercised on a different data value. Consequently, the first component is neither depth nor width bounded. However, any reachable configuration g will satisfy d∈D g(d)[2] = 1. In Petri net parlance, there is always exactly one token in the second place. The system def is thus place bounded for K = {2}. The main contribution of this paper is the effective computability of a suitable abstraction of the classical coverability tree construction [12] for UDPNs. This provides a way to decide all variants of the boundedness problem mentioned above. We summarise the consequences of our construction below. Theorem 2.5. In UDPNs, place depth boundedness implies place width boundedness. In consequence, place depth boundedness coincides with place boundedness. Theorem 2.6. The boundedness, place boundedness, and place width boundedness problems for UDPNs are in Fωω , i.e. in HyperAckermann. Let us emphasise the importance of decidability of place boundedness: first, the problem is undecidable in all the extensions of UDPNs in Fig. 1. Moreover, in the case of Petri nets, the decidability of place boundedness plays a crucial role in the decidability proofs for reachability [13, 14, 16, 18], hence Theorem 2.6 provides one of the basic building blocks for future attempts at proving the decidability of reachability for UDPNs.

3

Simple Ideals

A key observation about all decision problems mentioned in the previous section is that they do not require computing the reachability set: they can all be solved given some suitable representation of the cover [9], defined as def

Cover (f ) = ↓Reach(f ),

(4)

for f the initial configuration. Indeed, coverability reduces to checking whether g ∈ Cover (f ), boundedness to checking whether Cover (f ) is finite up to permutation of D, and place boundedness to checking whether {gK | g ∈ Cover (f )} is finite up to permutation of D. The main property of the coverability tree we construct in Sect. 4 is that we can extract a suitable representation of Cover (f ). Ideals and Clovers. We refer the reader to the work of Finkel and GoubaultLarrecq [8, 9] for details; it suffices to say that downwards-closed sets of configurations can be represented as finite unions of so-called configuration ideals. Formally, a configuration ideal J is a non-empty, downwards-closed, and directed

452

P. Hofman et al.

set of configurations; this last condition means that, if f and f  are configurations in J, then there exists h in J with f h and f  h. Crucially for algorithmic considerations, a configuration ideal J can in turn be represented as the downward-closure def (5) J = ↓g = {h ∈ Confs | h g} of a non-negative data vector g having a finite range: g(D) is a finite subset of Nkω . We can check that every such ↓g is a configuration ideal (see [8, 9] for the converse): it is non-empty and downwards-closed by definition, and we can check it is also directed. Indeed, if f , f  are configurations and π, π  are injections with f ≤ gπ and f  ≤ gπ  , then since f and f  are finitely supported we can assume π and π  to be permutations, and we can define a configuration h ≤ g such that f ≤ hπ and f  ≤ hπ  : set h as the pointwise least upper bound def −1 h(d)[i] = max(f (π −1 (d))[i], f  (π  (d))[i]) ≤ g(d)[i] for all d and 1 ≤ i ≤ k. Cover (f ), being downwards-closed, is represented by a finite set of representations of configuration ideals, called Clover (f ) by Finkel and Goubault-Larrecq:

Cover (f ) = {↓g | g ∈ Clover (f )} . Clover (f ) is determined uniquely up to permutation, and contains -maximal data vectors g satisfying ↓g ⊆ Cover (f ); for further details see [8, 9]. In the following we identify a configuration ideal J = ↓g with its representation g. Remark 3.1 (Ideals for Petri nets). For readers familiar with Karp and Miller’s coverability trees for Petri nets, observe that configuration ideal representations generalise the notion of ‘extended markings’, which are vectors in Nkω . Also, Clover (f ) for a Petri net can be computed as the set of vertex labels in its coverability tree—this will also be our case. Simple Ideals. Crucially, it turns out that we do not need general configuration ideals for our coverability trees for UDPNs. We only need to consider the downward-closures ↓g of non-negative vectors g (cf. (5)), where the set of vectors appearing infinitely often as g(d), when d ranges over D, is a singleton {I} for some vector I in {0, ω}k (instead of a finite subset of Nkω for general configuration ideals). Put differently, given such a vector I, we define the I-support of def

a data vector f as Supp I (f ) = {d ∈ D | f (d) = I}, and define an I-simple ideal (representation) as a non-negative data vector with finite I-support. In particular, a finitely supported non-negative data vector is a 0-simple ideal. We write M , N , . . . to denote simple ideals. A simple ideal M can be represented concretely as a pair M = m, I where m is the finite multiset of vectors in Nkω obtained from M by restriction to its I-support. Example 3.2. We represent simple ideals similarly as configurations, using the additional last column for the I part. Continuing with the UDPN of Example 2.2, its cover is the downward-closure of a single I-simple ideal:     Cover (f0 ) = ↓ ω1  ω0 , Clover (f0 ) = ω1  ω0 , def

where I = (ω, 0). The I-support of the ideal has one element, mapped to (ω, 1).

Coverability Trees for Petri Nets with Unordered Data

453

Note that UDPN steps map I-simple ideals to I-simple ideals. Lemma 3.3 formally states the relation between steps of ideals and steps of configurations in the downward closures. The next lemma shows that I-simple ideals can only have finitely many successors up to permutation. This property will later be used to define coverability trees of finite branching degree. → M  . Then for every Lemma 3.3. Let M, M  be I-simple ideals such that M −   configuration c ∈ ↓M there exist configurations c ∈ ↓M and c ∈ ↓M  with c− → c and c c . t

def

→ M  for a finite data vector t. The data vector f = c − t Proof. Suppose M − satisfies f ≤ M but f (d) can possibly be negative for some data value d; therefore def

we can not simply put c = f . A way to fix this is to define c by def

c(d)[i] = max(0, x(d)[i]),

for all d ∈ D and all coordinates i. def

Thus defined, c satisfies c ≤ M , and setting c = c + t satisfies c ≤ c as required.   Lemma 3.4. Let M be a simple ideal. The set {N | M − → N } of successors of M is finite up to permutation and has a representative with cardinality bounded by (|Supp I (M )| + maxt∈T |Supp 0 (t)|)! · |T |2 . Proof. Consider an UDPN defined by the finite set T of data vectors. Fix an Isimple ideal M , and denote by S the I-support of M . We will be now considering S-permutations, by which we mean those data permutations π that satisfy π(d) = d for all d ∈ S. Equality and finiteness up to S-permutation can be defined exactly as for plain permutations. A crucial but simple observation is that the set of transitions of the UDPN is finite up to S-permutation. Indeed, assume wlog. that S is disjoint from the supports of all vectors in T . Consider the finite set T  that contains all data vectors tσ, where t ∈ T and permutation σ swaps some subset of S with some subset of the support of t. Then every transition of the UDPN is of the form t π, where t ∈ T  and π is an S-permutation. Regarding the size of this new UDPN, there are at most t∈T (|S| + |Supp 0 (t)|)! such permutations σ, hence |T  | ≤ t∈T (|S| + |Supp 0 (t)|)! · |T |. Now we use the extension, to simple ideals, of the closure of the step relation under permutations, cf. Eq. (2), to derive a strengthening of our claim, namely finiteness of the successors of M up to S-permutations. Consider an arbitrary t π

step M −−→ N of M ; by Eq. (2) we get t

M = M π −1 − → N π −1 (the equality holds as π is an S-permutation). Therefore N is equal up to Spermutation to some T  -successor of M . As T  is finite, the set of T  -successors   of M is finite and bounded by |T  |, which implies our claim.

454

P. Hofman et al.

A consequence of our construction of coverability trees in Sect. 4, and of the complexity analysis conducted in Sect. 5, is the following core result: Theorem 3.5. Given an UDPN and an initial configuration f , an ideal representation Clover (f ) of Cover (f ) is computable in Fωω . Furthermore, Clover (f ) contains only simple ideals. Theorem 3.5, together with the following proposition, easily imply def Theorems 2.5 and 2.6 (below Clover (f )K = {gK | g ∈ Clover (f )}): Proposition 3.6. Fix K ⊆ {1, . . . , k}. An UDPN is width-bounded iff Clover (f )K contains only finitely supported vectors. An UDPN is depth-bounded iff Clover (f )K contains only finite vectors. Proof. The former equivalence, as well as the if direction of the latter one, follow by finiteness of Clover (f )K . It remains to argue that place depth-boundedness forces Clover (f )K to contain only finite vectors. Indeed, a non-finite simple ideal has necessarily ω at some component, which implies depth-unboundedness.   The remaining part of the paper is devoted to the proof of Theorem 3.5. In Sect. 4, we present an algorithmic construction of the coverability tree, and show its termination and correctness. Then in Sect. 5 we provide upper and lower bounds on the size of the coverability tree.

4

Representing a Cover

We will show that, analogously to the classical construction of Karp and Miller [12] for vector addition systems, the cover set of any UDPN configuration can be effectively represented in the form of a finite coverability tree, where nodes are labelled by simple ideals. For a given initial ideal the construction of a coverability tree amounts to iteratively computing successors (up to permutation), applying symbolic acceleration steps when a strictly dominating pair M  M  appears on a branch, and terminating a branch if a label embeds into one of its ancestors. 4.1

Accelerations

The idea behind acceleration steps is that due to monotonicity (Lemma 2.3), any finite sequence of steps t

t

t

1 2 k M1 −→ . . . −→ Mk M0 −→

(6)

such that M0  Mk can be extended indefinitely. Such an unfolding may have two distinct kinds of effect: Firstly, it may unboundedly increase components in data values already contained in the initial I-support (we call this effect depth acceleration). Secondly, it may increase an unbounded number of ‘fresh’ data values, outside of the initial I-support (we call this effect width acceleration).

Coverability Trees for Petri Nets with Unordered Data

455

Our construction only accelerates increasing sequences as above when there is a permutation (i.e. bijective) embedding of M0 into Mk . As a building block we shall use the usual vector acceleration: for two nonnegative vectors v, v  ∈ Nkω with v  ≤ v, define a new vector acc(v  , v), for 1 ≤ i ≤ k by:  v[i], if v  [i] = v[i], def  acc(v , v)[i] = ω, if v  [i] < v[i]. Definition 4.1 (Depth and Width Acceleration). For I-simple ideals M  , M and a permutation π with M  < M π, or equivalently M  π −1 < M , the depth acceleration of M  , M, π is the I-simple ideal defined by def

Mdepth (d) = acc(M  (π −1 (d)), M (d)),

for all data values d ∈ D. def

For d ∈ D such that M  (π −1 (d)) = I < M (d), put Id = acc(M  (π −1 (d)), M (d)); the width acceleration of M  , M, π, d is the Id -simple ideal defined by  if M (d) = I Id , def Mwidth (d) = M (d), otherwise, By definition, M < Mdepth , Mwidth . 4.2

Coverability Trees

By Lemma 3.4 we can compute for any I-simple ideal M a successor representative, namely a finite set such that every successor of M is equal up to permutation to some element of this set. For the sake of simplicity, we choose a conservative policy of application of accelerations: first, a proper nesting is imposed, in the sense that two different accelerated paths are either disjoint, or contained one in the other; second, a depth-accelerated path can not contain another accelerated path, while a widthaccelerated path can. However, as width accelerations strictly increase the I part, a width-accelerated path is never contained in another accelerated path. Therefore the only allowed inclusion is when a depth-accelerated path is included in a width-accelerated one. Definition 4.2 (Coverability Tree). A coverability tree is a tree with nodes labelled by simple ideals such that the following criteria are satisfied. 1. A node with label N is a leaf iff it has an ancestor with label N   N . 2. Otherwise, suppose an interior node N has an ancestor N  such that both N  , N are I-simple and N   N . Let P denote the path from N  to N in the tree, including N  and N . (a) Suppose N  (π −1 (d)) = I < N (d) for some permutation π with N  π < N and d ∈ D; and for every node in P that is a depth acceleration of some nodes M  , M , both M  and M belong to P. Then N has exactly one child labelled by the width acceleration of N  , N, π, d.

456

P. Hofman et al.

(b) Otherwise, if P contains no acceleration then N has exactly one child labelled by the depth acceleration of N  , N, π, for some permutation π with N  π < N . 3. Otherwise, if a node N satisfies none of the above criteria then its set of children is the successor representative of N . Remark 4.3. Note that Definition 4.2 does not determine the coverability tree unambiguously: the choice of a permutation π in points 2(a) and 2(b) is not unique. Remark 4.4. The condition in point 1 in Definition 4.2 implies that no branch of a coverability tree contains two different nodes with the same label. We identify a node with its label in the sequel. Example 4.5. We pick some coverability tree for the UDPN from Example 2.2 rooted in the configuration f0 . There is a branch with labels (up to permutation)             f0 = 11  00 , f1 = 21  00 , f2 = ω1  00 , f3 = ω0 11  00 , f4 = ω0 11  ω0 , f5 = ω0 11 01  ω0 , where f2 is a depth acceleration (of f0 , f1 ), f4 is a width acceleration (of f0 , f3 ), and all other nodes are the result of successor steps from their parent. The node f5 is a leaf because f5 f4 . Correctness. A coverability tree is finite (termination), and represents the cover of its root node (completeness and soundness). These required properties are proven in detail in the full paper: – Finiteness is proven by first exhibiting a wqo for the specific type of I-simple ideals that appears on coverability trees. This wqo depends on the existence of permutation embeddings, a property that on its own does not induce a wellquasi-ordering over the set of all I-simple ideals. Our termination argument is further refined to derive complexity bounds; see Sect. 5.2. – The completeness proof relies on the monotonicity of steps over simple ideals, and shows that all the elements in Cover (f ) are covered by some simple ideal in any coverability tree. – Soundness is the most delicate property to establish. Its crux is that neither width nor depth accelerations may take us outside the cover of the initial configuration.

5

Complexity Bounds

In the section, we prove lower and upper bounds on the resources needed by the construction of the coverability tree. We refer the reader to [24, 25] for gentle introductions to the techniques employed to prove these results. The enormous complexities involved in our construction require to use fast-growing complexity classes [23], which we present succinctly in Sect. 5.1 and in more details in the full paper, before showing hyper-Ackermannian upper and lower bounds in Sects. 5.2 and 5.3.

Coverability Trees for Petri Nets with Unordered Data

5.1

457

Fast-Growing Complexity

In order to express the non-elementary functions required for our complexity statements, we shall employ a family of subrecursive functions (hα )α indexed by ordinals α known as the Hardy hierarchy. Ordinal Terms. We use ordinal terms α in Cantor Normal Form (CNF), which can be written as terms α = ω α1 + · · · + ω αn where α1 ≥ · · · ≥ αn are themselves written in CNF. Using such notations, we can express any ordinal below ε0 , the minimal fixpoint of x = ω x . The ordinal 0 is obtained when n = 0; otherwise if αn = 0 the ordinal α is a successor ordinal ω α1 + · · · + ω αn−1 + 1, and if αn > 0 the ordinal α is a limit ordinal. We usually write λ to denote limit ordinals. Fundamental Sequences. For all x in N and limit ordinals λ, we use a standard assignment of fundamental sequences λ(0) < λ(1) < · · · < λ(x) < · · · < λ with supremum λ. Fundamental sequences are defined by transfinite induction by: (γ + ω β+1 )(x) = γ + ω β · (x + 1) , def





(γ + ω λ )(x) = γ + ω λ (x) . def

(7)

For instance, ω(x) = x + 1, ω 2 (x) = ω · (x + 1), ω ω (x) = ω x+1 , etc. The Hardy Hierarchy. Let h: N → N be a strictly increasing function. The Hardy functions (hα : N → N)α are defined by transfinite induction on their ordinal indices by def

h0 (x) = x ,

hα+1 (x) = hα (h(x)) , def

hλ (x) = hλ(x) (x) . def

(8)

k

Observe that h (x) for a finite k is simply the kth iterate of h. For limit ordinals def

λ, hλ (x) performs a form of diagonalisation: for instance, setting H(x) = x + 1 2 the successor function, H ω (x) = H x + 1 (x) = 2x + 1, H ω (x) = 2x + 1 (x + 1) − 1 3 is a function of exponential growth, while H ω is a non elementary function ω akin to a tower of exponentials of height x, H ω is a non primitive-recursive ωω is a non function with growth similar to the Ackermann function, and H ω multiply-recursive function characteristic of hyper-Ackermannian complexity. Complexity Classes. Following [23], we can define complexity classes for computations with time or space resources bounded by Hardy functions of the size of the input. We concentrate in this paper on the HyperAckermann complexity class. Let FMR denote the set of multiply-recursive functions and let h be any multiply-recursive strictly increasing function, then [23, Theorem 4.2]:

ωω def DTime(hω (m(n))) (9) HyperAckermann = Fωω = m∈FMR

is the set of decision problems solvable with resources bounded by an hyperAckermannian function applied to a multiply-recursive function m of the size of the input. This class is closed under multiply-recursive reductions, and several problems are known to be complete for it (see Sect. 6.2 of [23] for a survey), including coverability in unordered data nets [20].

458

P. Hofman et al.

5.2

Upper Bounds

We focus on the worst-case norm of the constructed simple ideals, from which bounds on the total size of the coverability tree and the complexity upper bound in Theorem 3.5 can both be derived. Norms of Simple Ideals. For a vector u in Zkω , its norm is its maximal finite def absolute value: v = max{|v[i]| | 1 ≤ i ≤ k ∧ v[i] = ω}. Observe that, k if I is in {0, ω} , then I = 0. For an I-simple ideal M , and thus for finitely supported ones in particular, we define its norm as the maximum between the cardinality of its support and the maximal norm of its vectors: def M  = max{|Supp I (M )|, M (d) | d ∈ D}. Note that the vectors for data d outside the support have all norm 0. In the full paper, we exhibit a bound def

ω k+3

(f0 ) on the norms of all the simple ideals constructed in a coverB = hω ability tree rooted by f0 as defined in Definition 4.2, where h also depends on the UDPN: Theorem 5.1. The norms of the simple ideals in a coverability tree rooted in a configuration f0 for a k-dimensional UDPN T are bounded by hω where h(x) is an elementary function of x, k, and T .

ω k+3

(f0 ),

The main technical ingredients for Theorem 5.1 are combinatorial statements on the lengths of so-called controlled bad sequences proven by Rosa-Velardo [20, Appendix A] for finite multisets of vectors of natural numbers. Our proofs require however a substantial amount of work on top of that of Rosa-Velardo’s for two reasons: we work with extended vectors in Nkω , and use permutation embeddings rather than just plain embeddings. Relating Norms with Sizes and Complexity. The norm M  ≤ B of a simple ideal M is directly related to the size of its concrete binary representation: the latter needs at most M  · k · (logM  + 1) bits for the I supported part of the ideal and k bits for the I vector itself. We can also bound the length of the branches in our coverability trees: there are indeed at most (B + 2)kB · 2k different simple ideals with norm ≤ B, and no two interior nodes on a branch are labelled by the same ideal due to condition 1 in Definition 4.2 (see Remark 4.4). Finally, by Lemma 3.4, the branching degree of the coverability tree is bounded by an exponential function (B + T )! · |T |2 in B and the size of T . These three observations combined allow to bound the size of the coverability tree: Theorem 5.2 (Size of Coverability Trees). The size of a coverability tree built from an initial configuration f0 for a k-dimensional UDPN T is bounded by an elementary function of B, k, and the size of T . Theorem 5.2 along with Eq. (9) and the completeness and soundness of coverability trees yields the proof of Theorem 3.5, using the fact that Fωω is closed under elementary reductions [23, Theorem 4.7].

Coverability Trees for Petri Nets with Unordered Data

5.3

459

Lower Bounds

The sheer complexity bounds we just obtained on the size of coverability trees beg the question whether they are the best possible. We show in Theorem 5.3 that, indeed, the size of coverability trees for a family of UDPNs is provably non multiply-recursive, matching essentially the statement of Theorem 5.2: Theorem 5.3 (Hyper-Ackermannian Coverability Trees). There exists families of O(k)-sized UDPNs (Tk )k and O(k + log n)-sized initial configurations (fk,n )k,n , whose coverability trees are of size at least H ω

ωk

(n).

Hardy Computations. As detailed in the full paper, we prove Theorem 5.3 by ωk

‘implementing’ the computation of Hardy functions H ω in nets Tk . The main idea, first developed in [11, 24], is to see the equations in (8) for 0 < α as rewriting rules operating on pairs (α, n): (α + 1, n) → (α, n + 1) ,

(λ, n) → (λ(n), n) .

(10)

Note that a sequence (α0 , n0 ) → (α1 , n1 ) → · · · → (αi , ni ) → · · · of rewriting steps maintains H αi (ni ) = H α0 (n0 ) for all i, and must eventually terminate at some rank with α = 0 since αi > αi+1 for all i, and then n = H α0 (n0 ). k Using a natural representation of ordinals α < ω ω as finite multisets of vectors also employed by Rosa-Velardo [20], a pair (α, n) can be encoded as a configuration of Tk , and the rewriting rules of (10) can be implemented on such codes by steps of Tk . This is however not a perfect implementation: many incorωk

ω k−1 ·(n+1)

(n) rect computations yielding results different from H ω (n) = H ω are possible. The crucial point is that there exists a perfect computation in Tk , ωk

of length at least H ω (n). Furthermore, this computation does not allow any acceleration step, and has therefore to occur as such in any coverability tree.

6

Concluding Remarks

In this paper, we have presented a procedure to construct coverability trees for UDPNs in the style of Karp and Miller [12]. This yields decision procedures for coverability and several variants of the boundedness problem including placeboundedness, depth- and width place-boundedness. Besides its interest for the formal verification of UDPNs, this paves the way towards future attempts at proving the decidability of reachability along the lines developed for Petri nets in [13,14,16, 18]. We have derived hyper-Ackermannian upper bounds on the complexity of our construction, and shown that such enormous complexities are actually attained on some UDPNs. Note that this however does not provide a lower bound – on the size of Clover (f ), for which the best known bound is an Ackermannian lower bound adapted from the case of Petri nets [4], nor

460

P. Hofman et al.

– on the complexity of the various boundedness problems on UDPNs, for which the best lower bound is hardness for Tower = F3 , adapted from the coverability problem [15]. We actually suspect that much lower complexities that HyperAckermann could be obtained for the coverability and boundedness problems. For instance, in the case of Petri nets, coverability trees have a worst-case Ackermannian size [4, 6], but coverability, boundedness, and place-boundedness are all ExpSpace-complete [2, 5,17, 19].

References ˇ ans, K., Jonsson, B., Tsay, Y.K.: Algorithmic analysis of pro1. Abdulla, P.A., Cer¯ grams with well quasi-ordered domains. Inform. and Comput. 160(1–2), 109–127 (2000) 2. Blockelet, M., Schmitz, S.: Model checking coverability graphs of vector addition systems. In: Murlak, F., Sankowski, P. (eds.) MFCS 2011. LNCS, vol. 6907, pp. 108–119. Springer, Heidelberg (2011) 3. Boja´ nczyk, M., Klin, B., Lasota, S.: Automata theory in nominal sets. Logic. Meth. Comput. Sci. 10(3:4), 1–44 (2014) 4. Cardoza, E., Lipton, R.J., Meyer, A.R.: Exponential space complete problems for Petri nets and commutative semigroups: preliminary report. In: STOC 1976, pp. 50–54. ACM (1976) 5. Demri, S.: On selective unboundedness of VASS. J. Comput. Syst. Sci. 79(5), 689– 713 (2013) 6. Figueira, D., Figueira, S., Schmitz, S., Schnoebelen, P.: Ackermannian and primitive-recursive bounds with Dickson’s Lemma. In: LICS 2011, pp. 269–278. IEEE Press (2011) 7. Finkel, A., Schnoebelen, P.: Well-structured transition systems everywhere!. Theor. Comput. Sci. 256(1–2), 63–92 (2001) 8. Finkel, A., Goubault-Larrecq, J.: Forward analysis for WSTS, part I: completions. In: STACS 2009. LIPIcs, vol. 3, pp. 433–444. LZI (2009) 9. Finkel, A., Goubault-Larrecq, J.: Forward analysis for WSTS, part II: complete WSTS. Logic. Meth. Comput. Sci. 8(3:28), 1–35 (2012) 10. Finkel, A., McKenzie, P., Picaronny, C.: A well-structured framework for analysing Petri net extensions. Inform. and Comput. 195(1–2), 1–29 (2004) 11. Haddad, S., Schmitz, S., Schnoebelen, P.: The ordinal recursive complexity of timed-arc Petri nets, data nets, and other enriched nets. In: LICS 2012, pp. 355– 364. IEEE Press (2012) 12. Karp, R.M., Miller, R.E.: Parallel program schemata. J. Comput. Syst. Sci. 3(2), 147–195 (1969) 13. Kosaraju, S.R.: Decidability of reachability in vector addition systems. In: Proceedings STOC 1982, pp. 267–281. ACM (1982) 14. Lambert, J.L.: A structure to decide reachability in Petri nets. Theor. Comput. Sci. 99(1), 79–104 (1992) 15. Lazi´c, R., Newcomb, T., Ouaknine, J., Roscoe, A., Worrell, J.: Nets with tokens which carry data. Fund. Inform. 88(3), 251–274 (2008) 16. Leroux, J., Schmitz, S.: Demystifying reachability in vector addition systems. In: LICS 2015, pp. 56–67. IEEE Press (2015)

Coverability Trees for Petri Nets with Unordered Data

461

17. Lipton, R.: The reachability problem requires exponential space. Technical Report 62, Yale University (1976) 18. Mayr, E.W.: An algorithm for the general Petri net reachability problem. In: Proceedings STOC 1981, pp. 238–246. ACM (1981) 19. Rackoff, C.: The covering and boundedness problems for vector addition systems. Theor. Comput. Sci. 6(2), 223–231 (1978) 20. Rosa-Velardo, F.: Ordinal recursive complexity of unordered data nets. Technical Report TR-4-14, Departamento de Sistemas Inform´ aticos y Computaci´ on, Universidad Complutense de Madrid (2014). http://antares.sip.ucm.es/frosa/docs/ complexityUDN.pdf 21. Rosa-Velardo, F., de Frutos-Escrig, D.: Decidability and complexity of Petri nets with unordered data. Theor. Comput. Sci. 412(34), 4439–4451 (2011) 22. Rosa-Velardo, F., Martos-Salgado, M., de Frutos-Escrig, D.: Accelerations for the coverability set of Petri nets with names. Fund. Inform. 113(3–4), 313–341 (2011) 23. Schmitz, S.: Complexity hierarchies beyond elementary. ACM Trans. Comput. Theor. (2016) (to appear). http://arxiv.org/abs/1312.5686 24. Schmitz, S., Schnoebelen, P.: Algorithmic aspects of WQO theory. Lecture notes (2012).http://cel.archives-ouvertes.fr/cel-00727025 25. Schmitz, S., Schnoebelen, P.: The power of well-structured systems. In: D’Argenio, P.R., Melgratti, H. (eds.) CONCUR 2013 – Concurrency Theory. LNCS, vol. 8052, pp. 5–24. Springer, Heidelberg (2013) 26. Schnoebelen, P.: Revisiting Ackermann-hardness for lossy counter machines and reset Petri nets. In: Hlinˇen´ y, P., Kuˇcera, A. (eds.) MFCS 2010. LNCS, vol. 6281, pp. 616–628. Springer, Heidelberg (2010)

Shortest Paths in One-Counter Systems nski2 , Piotr Hofman3 , Dmitry Chistikov1(B) , Wojciech Czerwi´ 2 Michal Pilipczuk , and Michael Wehar4 1

2

Max Planck Institute for Software Systems (MPI-SWS), Kaiserslautern and Saarbr¨ ucken, Germany [email protected] Institute of Informatics, University of Warsaw, Warsaw, Poland {wczerwin,michal.pilipczuk}@mimuw.edu.pl 3 Laboratoire Sp´ecification et V´erification (LSV), ENS Cachan & CNRS, Paris, France [email protected] 4 Department of Computer Science and Engineering, University at Buffalo, Buffalo, USA [email protected]

Abstract. We show that any one-counter automaton with n states, if its language is non-empty, accepts some word of length at most O(n2 ). This closes the gap between the previously known upper bound of O(n3 ) and lower bound of Ω(n2 ). More generally, we prove a tight upper bound on the length of shortest paths between arbitrary configurations in onecounter transition systems (weaker bounds have previously appeared in the literature).

1

Introduction

Extremal combinatorial questions are ubiquitous in today’s theory of computing: How many steps does an algorithm take in the worst case when traversing a data structure? How large is the most compact automaton for a formal language? While some specific questions of this form are best seen as standalone puzzles, only interesting for their own sake, others can be used as basic building blocks for more involved arguments. We look into the following extremal problem: Given a one-counter automaton A with n states, how long can the shortest word accepted by A be? It is folklore that, unless the language of A is empty, A accepts some word of length at most polynomial in n. This fact and a number of related results of similar form have The main part of these results was obtained during Autob´ oz’15, the annual research camp of Warsaw automata group. During the work on these results, Mi. Pilipczuk held a post-doc position at Warsaw Centre of Mathematics and Computer Science and was supported by the Foundation for Polish Science via the START stipend programme. P. Hofman was supported by Labex Digicosme, Univ. Paris-Saclay, project VERICONISS. W. Czerwi´ nski acknowledges a partial support by the Polish National Science Centre grant 2013/09/B/ST6/01575. c Springer-Verlag Berlin Heidelberg 2016  B. Jacobs and C. L¨ oding (Eds.): FOSSACS 2016, LNCS 9634, pp. 462–478, 2016. DOI: 10.1007/978-3-662-49630-5 27

Shortest Paths in One-Counter Systems

463

appeared as auxiliary lemmas in the literature on formal languages, analysis of infinite-state systems, and applications of formal methods [12, Lemma 6], [11, Sect. 8.1], [8, Lemma 5], [1, Lemma 11], [9, Lemmas 28 and 29], [7, Sect. 5]. A closer inspection reveals that the arguments behind these results deliver (or can deliver) an upper bound of O(n3 ), while the best known lower bound comes from examples of one-counter automata with shortest accepted words of length Θ(n2 ). In other words, the true value is at least quadratic and at most cubic. The main result of this paper is that we close this gap by showing a quadratic upper bound, O(n2 ). We also extend this result to a more general reachability setting: in any one-counter (transition) system with n control states, whenever there is a path from a configuration α to a configuration β—recall that configurations are pairs of the form (q, c) where q ∈ Q is the control state, |Q| = n, and c is a counter value, a nonnegative integer—there is also a path from α to β that has length at most O(n2 + n · max(cα , cβ )) where cα and cβ are the counter values of α and β. We discuss our contribution in more detail in Sect. 2. Related Work and Motivation. Reachability is a fundamental problem in theoretical computer science and in its applications in verification, notably via analysis of infinite-state systems [2, 4,14, 17]. Among such systems, counter-based models of computation are a standard abstraction that has attracted a lot of attention [3]; machines with a single counter are, of course, the most basic. Nevertheless, while our main motivation has been purely theoretical, we note that bounds on the length of shortest paths in one-counter systems have appeared as building blocks in the literature on rather diverse topics. More specifically, a polynomial upper bound is used by Etessami et al. [8] in an analysis of probabilistic one-counter systems (which are equivalent to socalled discrete-time quasi-birth-death processes, QBDs). They prove that in the (q, 1)  (q  , 0)-reachability setting the counter does not need to grow higher than n2 and provide examples showing that this bound is tight. However, they only deduce upper bounds of n3 and n4 on the length of shortest paths without and with zero tests, respectively. A simple corollary shows that if a state q can eventually reach a state q  with a non-zero probability, then this probability is lower-bounded by p poly(n) where p is the smallest among positive probabilities associated with transitions. This becomes a step in the proof that a (decomposed) Newton’s method approximates termination probabilities of the system in time polynomial in its size, n; the results of the present paper reduce the (theoretical) worst-case upper bounds on the number of steps roughly by a factor of n. In a subsequent work, Hofman et al. [9] reuse the auxiliary lemmas on the length of shortest paths from [8] and show that (strong and weak) trace inclusion for a one-counter system and a finite-state process is decidable in PSPACE (and is, in fact, PSPACE-complete). One may note that a stronger upper bound of O(n3 ) on the length of shortest paths can be derived from the above bound on the largest needed counter value even in the presence of zero tests. This value, O(n3 ), seems to be a recurring theme in the literature on one-counter systems; it already appears in the pumping lemma for one-counter languages due to Latteux [13] as the pumping constant: a

464

D. Chistikov et al.

number N such that any accepted word longer than N can be pumped. In fact, the formulation in [13] does not permit removals of factors from an accepted word, but even such a version would only yield the same upper bound of O(n3 ) on the length of shortest paths. While the arguments of the present paper do not lead to an improvement in the pumping constant for one-counter languages, we nevertheless show that in the reachability setting the optimal value (the length of the shortest path) is actually O(n2 ). A cubic upper bound on the largest needed counter value (for the reachability setting) in one-counter systems without zero tests, also known as one-counter nets, appears in the work of Lafourcade et al. [11, 12]. This result is applied in the context of the Dolev-Yao intruder model, where the question of whether a passive eavesdropper (an intruder) can obtain a piece of information is reduced to the decision problem for a deduction system. For several such systems, Lafourcade et al. show that, under certain assumptions, the problem is decidable in polynomial time. They construct a one-counter system where states represent terms from a finite set and the counter value corresponds to the number of applications of a free unary function symbol to a term. After this, the upper bound on counter values along shortest paths is extended to an upper bound on the size of terms that can be used in a minimal deductive proof; needless to say, an improvement in the upper bound extends in a natural way. ˇ Finally, we would like to mention the work of Alur and Cern´ y [1], who use a related model of one-counter systems with counter values in Z and without zero tests. They reduce the equivalence problem for so-called streaming datastring transducers to (q, 0)  (q  , 0)-reachability in such counter systems: the transducers produce output at the end of the computation, and the counter is used to track the accumulated distance between a distinguished pair of symbols in the output. Since these transducers are designed to model list-manipulating programs (in two syntactically restricted models), decision procedures for equivalence of such programs can rely on the upper bounds for shortest paths to efficiently prune the search space. In [1], the upper bound on the path length is the familiar O(n3 ); this gives an upper bound on the length of smallest counterexamples to equivalence. Our upper bound of O(n2 ) extends to this model of counter systems too; because of space constraints, details are only given in the full version of the paper. The reduction to reachability in one-counter systems was recently implemented by Thakkar et al. [16] on top of ARMC, an abstraction-refinement model checker [15], for the purpose of verifying retransmission protocols over noisy channels.

2

Summary

One-Counter Systems. In this paper we work in the framework of one-counter systems, which are an abstract version of one-counter automata. More precisely, they are one-counter automata without input alphabet (see below). Formally, a one-counter system (OCS) O consists of a finite set of states Q, a set of non-zero transitions T>0 ⊆ Q × {−1, 0, 1} × Q, and a set of zero tests

Shortest Paths in One-Counter Systems

465

T=0 ⊆ Q×{0, 1}×Q. A configuration of the OCS O is a pair in Q×N. We define a binary relation −→ on the set Q×N as follows: (p, c) −→ (q, c+d) whenever (i) c ≥ 1 and (p, d, q) ∈ T>0 or (ii) c = 0 and (p, d, q) ∈ T=0 . The reflexive transitive closure of −→ is denoted by −→ ∗ ; we say that a configuration β is reachable from α if α −→ ∗ β. This reachability is witnessed by a path in OCS O, which is simply a path in the infinite directed graph with vertices Q×N and edge relation −→ ; vertices and edges along the path can be repeated. The length of the path is the number of (not necessarily distinct) edges that occur on it. Our Contribution. We first formulate our results in terms of one-counter systems. Our first result is on paths between configurations with zero counter values. Theorem 1. Let O be a one-counter system with n states. Suppose a configuration β = (pβ , 0) is reachable from a configuration α = (pα , 0) in O. Then O has a path from α to β of length at most 14n2 . We then generalize the result to arbitrary source and target configurations. Theorem 2. Let O be a one-counter system with n states. Suppose a configuration β = (pβ , cβ ) is reachable from a configuration α = (pα , cα ) in O. Then O has a path from α to β of length at most 14n2 + n · max(cα , cβ ). The proof of Theorem 1 is the main technical contribution of this work. For this reason, in this extended abstract we focus on proving Theorem 1, while the reasoning leading to Theorem 2, as well as an extension to OCS with negative counter values, can be found in the full version of the paper. The full version contains also the proofs of all the statements marked with ♠. One-Counter Automata. We now restate our contribution in terms of onecounter automata (which are the original motivation for this work). Take any finite set Σ. The set of all finite words over Σ is denoted by Σ∗ , and the empty word by ε. A (nondeterministic) one-counter automaton A over Σ is a one-counter system where every transition t ∈ T>0 ∪ T=0 is associated with a label, λ(t) ∈ Σ ∪ {ε}, and where some subsets I ⊆ Q and F ⊆ Q are distinguished as sets of initial and final states respectively. The labeling function λ is extended from transitions to edges −→ and to paths in a natural way; the automaton accepts all words that are labels of paths from I × {0} to F × N. The language of a one-counter automaton A is the set of all words accepted by A. Corollary 1 (♠). Let A be a nondeterministic one-counter automaton with n states. If the language of A is non-empty, then A accepts some word of length at most 14n2 . As a concrete example, from Corollary 1 it follows that any nondeterministic one-counter automaton that accepts the singleton unary language {an } —a basic √ version of counting to n— must have at least Ω( n) states. This lower bound is tight and shows that nondeterminism does not help to “count √ to n”, because deterministic one-counter automata can also do this using Θ( n) states [5].

466

D. Chistikov et al.

Lower Bounds. As we already said, the lower bound on the length of the shortest path is Ω(n2 ). We present constructions of OCS that match the upper bounds of Theorems 1 and 2. Note that Examples 1 and 2 seem to use different phenomena. Example 1. [5,8] Consider an OCS O1 with 2n states: p1 , . . . , pn and q1 , . . . , qn . Let O1 have, for 1 ≤ i < n, transitions (pi , +1, pi+1 ) and (qi , 0, qi+1 ), as well as (qn , −1, q1 ) and (pn , 0, q1 ). All the transitions are non-zero, except for transition (p1 , +1, p2 ), which is a zero test. This OCS is deterministic: every configuration has at most one outgoing transition. The only path from (p1 , 0) to (q1 , 0) has length n2 . Example 2. [8] Let k and m be coprime and let OCS O2 have states p0 , . . . , pk−1 , q0 , . . . , qm−1 , and s1 , s2 . Let O2 have, for all 0 ≤ i < k and 0 ≤ j < m, non-zero transitions (pi , +1, pi+1 mod k ) and (qj , −1, qj+1 mod m ), a non-zero (p0 , −1, q1 ), and zero tests (s1 , +1, p1 ), (q0 , 0, s2 ). Now paths from (s1 , 0) to (s2 , 0) correspond to solutions of x·k −y ·m = 0; the shortest path takes the first cycle x = m times and the second cycle y = k times. Exiting the second cycle uses an additional transition, making the length 2km + 1. Setting k = n and m = n − 1 gives an OCS O2 with 2n + 1 states where not only does the shortest path have quadratic length, but all such paths also need to use quadratic counter values. Example 3. This example justifies the need for the term n · max(cα , cβ ) in Theorem 2. Modify O1 from Example 1 as follows. Add states a1 , . . . , an , b1 , . . . , bn and the following non-zero transitions: (an , −1, a1 ), (bn , +1, b1 ), and, for all 0 ≤ i < n, (ai , 0, ai+1 ) and (bi , 0, bi+1 ). For each of these non-zero transition, apart from (an , −1, a1 ), introduce also the same transition as a zero test. Finally, add two more zero tests: (an , 0, p1 ) and (q1 , 0, b1 ). Thus, the obtained OCS O3 has 4n states. Observe that every path in O3 from (a1 , cα ) to (bn , cβ ) has to go through (an , 0) and (b1 , 0) and thus has length at least n2 + n(cα + cβ + 2).

3

Challenges and Techniques

We now discuss shortly the intuition behind our approach to proving Theorem 1, and where the main challenges lie. The first, obvious observation is as follows: if some configuration appears more than once on a path, then the segment between any two appearances of this configuration can safely be removed. If we apply this modification exhaustively, then on each “level” — a set of configurations with the same counter value — we cannot see more than n configurations. If the maximum counter value observed on some path were bounded by O(n), then we would immediately obtain a quadratic upper bound on its length. Unfortunately, this is not the case: as Example 2 shows, the counter values in the shortest accepting path can be as large as quadratic. Hence, applying this observation in a straightforward manner cannot lead to any upper bound better than cubic. Instead, we perform an involved surgery on the path. The first idea is to start with a path ρ◦ that is not the shortest, but uses the fewest zero tests;

Shortest Paths in One-Counter Systems

467

the observation above shows that their number is bounded by n. Each subpath between two consecutive zero tests is called an arc, and we aim at modifying each arc separately to make it short. An arc is called low if it contains only configurations with counter values at most 5n, and high otherwise. The total length of low arcs can again be bounded by O(n2 ) by just excluding repeated configurations, so it suffices to focus on high arcs. Suppose ρ is a high arc. Since we observe high counter values on ρ, one can easily find a positive cycle σ + in the early parts of ρ, and a negative cycle σ − in the late parts of ρ. Here by a cycle we mean a sequence of transitions that starts and ends in the same state, and the cycle is positive/negative if the total effect it has on the counter during its traversal is positive/negative. Let A be the (positive) effect of σ + on the counter, and −B be the (negative) effect of σ − . Now comes the crucial idea of the proof: we can modify ρ by pumping σ + and − σ up many times, thus effectively “lifting” the central part of the path (called cap) to counter levels where there is no threat of hitting counter value zero while performing modifications (see Fig. 1, p. 11). More importantly, the cap can now be unpumped “modulo gcd(A, B)” in the following sense: we can exhaustively remove subpaths between configurations that have the same state and whose counter values are congruent modulo gcd(A, B). The reason is that any change in the total effect of the cap on the counter that is divisible by gcd(A, B) can be compensated by adjusting the number of times we pump cycles σ + and σ − . In particular, the length of the cap becomes reduced to at most gcd(A, B) · n, at the cost of pumping σ + and σ − several times. By performing this operation (we call it normalization) on all high arcs, we make them normal. After this, we apply an involved amortization scheme to show that the total length of normal arcs is at most quadratic in n. This requires very delicate arguments for bounding (i ) the total length of the caps and (ii ) the total length of the pumped cycles σ + and σ − throughout all the normal arcs. In particular, for this part of the proof to work we need to assert a number of technical properties of normal arcs; we ensure that these properties hold when we perform the normalization. Most importantly, whenever for two arcs the corresponding cycles σ + (or σ − ) lie in the same strongly connected component of the system (looking at the graph induced only by non-zero transitions), we stipulate that in both arcs σ + (or σ − ) actually refer to the same cycle. The final amortization is based on the analysis of pairs of strongly connected components to which σ + and σ − belong, for all normal arcs. At least as of now, arguments of this flavor (inspired by amortized analysis reasoning) are not typical for formal language theory and are more characteristic of the body of work on algorithms and data structures; see, e.g., [6, 10].

4

Preliminaries

In this paper N stands for the set of nonnegative integers. For any set X and a word w ∈ X ∗ , the length of w = x1 . . . xn , denoted len(w), is the number n of symbols in w. For k ∈ N and a word w, by wk we denote the word w repeated

468

D. Chistikov et al.

k times. For two positive integers x, y, by gcd(x, y) and lcm(x, y) we denote the greatest common divisor and the least common multiple of x and y, respectively. Recall that x · y = gcd(x, y) · lcm(x, y). We now give all definitions related to one-counter systems that we need later. For the reader’s convenience, concepts from Sect. 2 are defined anew. A one-counter system (OCS) O consists of a finite set of states Q, a set of non-zero transitions T>0 ⊆ Q × {−1, 0, 1} × Q, and a set of zero tests T=0 ⊆ Q × {0, 1} × Q. The set of transitions is T = T>0 ∪ T=0 . For a transition t = (q, d, q  ) ∈ T , by src(t) and targ(t) we denote q and q  , i.e., the source and the target state of t respectively. Further, the effect of the transition t = (q, d, q  ) is the number d; we write eff(t) m = d. We extend this notion to sequences of transitions: eff(t1 . . . tm ) = i=1 eff(ti ). A configuration of the OCS O is a pair in Q × N. The state of a configuration (q, c) is the state q; we also say that configuration (q, c) has state q, and write st((q, c)) = q. The counter value of configuration (q, c) is the number c; we write cnt((q, c)) = c. A transition t = (q, d, q  ) ∈ T can be fired in a configuration γ = (q, c) if either t ∈ T>0 and c > 0 or t ∈ T=0 and c = 0. In other words, zero tests can be fired only if the counter value is zero, and non-zero transitions can be fired only if the counter value is positive. The result of firing (q, d, q  ) in (q, c) is the t configuration γ  = (q  , c + d). We then write γ −→ γ  . A path ρ of the OCS O is a sequence of pairs (γ1 , t1 )(γ2 , t2 ) . . . (γm , tm ) ∈ ((Q × N) × T )∗ t

i such that for every i ∈ {1, . . . , m−1} we have γi −→ γi+1 and there exists a contm figuration γm+1 such that γm −→ γm+1 . The length of this path is m. The source of ρ, denoted by src(ρ), is γ1 ; we also say that ρ starts in its source. Similarly, the target of ρ, denoted by targ(ρ), is γm+1 ; we say that ρ finishes in its target. Note that now the source and target are configurations, rather than individual states; the path is from its source to its target. All γ2 , . . . , γm are called intermediate configurations. We also say that configurations γ1 , γ2 , . . . , γm+1 appear on ρ; note that the target of ρ also appears on ρ. Finally, when such a path exists, the configuration γm+1 is said to be reachable from the configuration γ1 . The projection of a path ρ is the sequence of its transitions t1 t2 . . . tm ; we write proj(ρ) = t1 t2 . . . tm . We follow the convention of denoting paths by ρ and sequences of transitions by σ. The effect of a path ρ is eff(ρ) = eff(proj(ρ)). A sequence of transitions σ = t1 t2 . . . tm is fireable in a configuration γ1 if there exists a path ρ = (γ1 , t1 )(γ2 , t2 ) . . . (γm , tm ). This path ρ is called the fastening of σ at γ1 , denoted ρ = fasten(γ1 , σ). Note that in particular proj(fasten(γ, σ)) = σ for every γ in which σ is fireable. A sequence of transitions t1 t2 . . . tm is consistent if for all i ∈ {1, . . . , m−1} it holds that targ(ti ) = src(ti+1 ). Note that a sequence of transitions fireable in some configuration is always consistent, but the other implication does not hold in general. We extend the notation src(·) and targ(·) to consistent sequences of transitions: src(t1 t2 . . . tm ) = src(t1 ) and targ(t1 t2 . . . tm ) = targ(tm ). The sources and targets of the transitions of t1 t2 . . . tm are visited on t1 t2 . . . tm .

Shortest Paths in One-Counter Systems

469

A cycle σ is a consistent sequence of non-zero transitions that starts and finishes in the same state q. This q is called the base state of the cycle σ. If the effect of σ is positive (resp. negative), then it is a positive (resp. negative) cycle. A cycle σ is called simple if every state is visited at most once on σ, except for the base of σ, which is visited only at the start and at the end.

5 5.1

Proof of Theorem 1 Proof Overview and Notation

Let us fix the OCS O we work with; let Q be its state set and let n = |Q|. Suppose ρ0 is a path from α to β, and let ρ0 be chosen such that it has the smallest possible number of configurations with counter value zero. Note that ρ0 does not have to be the shortest path between α and β. The first step is to divide ρ0 into subpaths, called arcs, between consecutive configurations with counter value zero. Then we modify the arcs separately. If a counter value in an arc does not exceed 5n, then we say that the arc is low, otherwise it is high. The low arcs will not be changed at all, and the reason is that we can bound quadratically the total number of configurations with counter value at most 5n using the following straightforward proposition. It is similar, in the spirit, to pumping lemmas, but simply removes a part of the path. Proposition 1. Suppose ρ = (γ1 , t1 )(γ2 , t2 ) . . . (γm , tm ) is a path from α to β. Suppose further that for some i and j with 1 ≤ i < j ≤ m + 1 it holds that tm γm+1 . Consider γi = γj , where γm+1 is such that γm −→ ρ = (γ1 , t1 )(γ2 , t2 ) . . . (γi−1 , ti−1 )(γj , tj )(γj+1 , tj+1 ) . . . (γm , tm ). Then ρ is also a path from α to β. However, the high arcs will be heavily modified. Roughly speaking, if an arc is high, then it contains both a positive cycle near its beginning and a negative cycle near its end. We can use these cycles to pump the middle part of the path as much up as we like. Thus, the modified path will consist of a short prefix; then several iterations of a positive cycle pumping it up; then a so called cap: a part of the path with only high counter values; then several iterations of a negative cycle pumping it down; and finally a short suffix. We show in the sequel how to perform this construction in such a way that the total length of pumping cycles, short prefixes and suffixes, and caps is quadratic. The construction itself (with arc-level length estimates) is presented in the following Subsect. 5.2, and the upper bound on the length of the entire path is given in Subsect. 5.3. Transition multigraph. One can view a transition (p, c, q) ∈ Q×{−1, 0, 1}×Q also as an edge (p, q) ∈ Q × Q labelled by a number c ∈ {−1, 0, 1}. In the proof we will many times switch back and forth between these two perspectives. In order to keep the mathematical precision we introduce a bit of notation.

470

D. Chistikov et al.

The transition multigraph G = (V, E, ) of an OCS consists of a set of nodes V , a multiset of directed edges E, and a labeling : E → {−1, 0, 1}. Set V equals the set of states Q. Every non-zero transition t = (p, c, q) ∈ T>0 in O gives rise to an edge e = (u, v) ∈ E with (e) = c. Note that the definition of the transition multigraph does not take into account the zero transitions. In the proof we pay a special attention to strongly connected components (SCCs) of G. Recall that two vertices p, q ∈ V are said to communicate if G has a walk from p to q and a walk from q to p. Communication is an equivalence relation, and its equivalence classes are called the strongly connected components of G. Let S be the set of all strongly connected components of G. For a strongly connected component S ∈ S, by nS we denote the number of vertices in S. We say that a cycle σ is contained in S if each state appearing on σ belongs to S. Note that every cycle is contained in some SCC, and a simple cycle contained in S has length at most nS . We say that an SCC S is positively enabled if it contains a cycle that has a positive effect. Similarly, S is negatively enabled if it contains a cycle that has a negative effect. Note that an SCC S can be both positively and negatively enabled. Lemma 1 (♠). Let G be a transition multigraph of an OCS and S a positively (respectively, negatively) enabled SCC. Then there exists a positive (respectively, negative) cycle σ contained in S that is simple. For every positively enabled SCC S we distinguish one, arbitrarily chosen, simple cycle with positive effect contained in S; we denote it by σS+ . Its existence is guaranteed by Lemma 1. Similarly, for every negatively enabled S we distinguish one simple cycle with negative effect contained in S, and we denote it by σS− . The base states of these cycles are chosen arbitrarily. 5.2

Normal Paths

A path is an arc if both its source and target have counter value zero, but all its intermediate configurations have counter values strictly larger than zero. An arc (or a path) is low if all its configurations (including the target) have counter values strictly smaller than 5n. An arc ρ is (S, T )-normal, where S and T are some SCCs of the transition multigraph, if it admits the following normal decomposition (see Fig. 1, p. 11): ρ = ρpref ρup ρcap ρdown ρsuff , where – – – – –

ρpref and ρsuff are low; proj(ρup ) = (σup )k for some k ∈ N, where σup = σS+ ; proj(ρdown ) = (σdown ) for some ∈ N, where σdown = σT− ; st(src(ρcap )) is the base state of σup ; and st(targ(ρcap )) is the base state of σdown .

Shortest Paths in One-Counter Systems

471

We say that an arc ρ is normal if it is (S, T )-normal for some S, T ∈ S. See Fig. 1 for an illustration. Then a path ρ is normal if it is a concatenation of normal arcs (possibly for different pairs (S, T )) and low arcs. In the remaining part of the proof we will show that if β is reachable from α, where cnt(α) = cnt(β) = 0, then there exists a short normal path from α to β. We start by analyzing a single arc. The following lemma, which is the most technically involved step in this paper, shows that we can restrict ourselves to normal arcs that have a very special structure. Lemma 2 (♠). If cnt(α) = cnt(β) = 0 and there exists an arc from α to β, then there exists an arc ρ from α to β which is either low or normal. Moreover, in the case when ρ is normal, a normal decomposition ρ = ρpref ρup ρcap ρdown ρsuff can be chosen such that: (i) proj(ρup ) = (σup )a , eff(σup ) = A for some a, A ∈ N; (ii) proj(ρdown ) = (σdown )b , eff(σdown ) = −B for some b, B ∈ N; (iii) a · A ≤ 2 · len(ρcap ) + 2 · lcm(A, B); (iv) b · B ≤ 2 · len(ρcap ) + 2 · lcm(A, B); (v) no infix of proj(ρcap ) is a cycle with effect divisible by gcd(A, B); (vi) cnt(targ(ρup )), cnt(src(ρdown )) > n; and (vii) all configurations appearing on ρpref and ρsuff are pairwise different. We now explain some intuition behind this statement. First note that, by condition (vii), the total number of configurations appearing on ρpref and ρsuff is at most 5n · n, since n is the number of states of the OCS O and both of these paths are low (so counter values 5n and above do not occur). Thus, len(ρpref ) + len(ρsuff ) ≤ 5n2 . Second, we can conclude from condition (v) that every state q ∈ Q can occur in configurations appearing in ρcap at most gcd(A, B) times; hence, len(ρcap ) ≤ n · gcd(A, B) ≤ n2 . Finally, condition (i) implies len(ρup ) ≤ a · n; if, for instance, a ≤ const · n, then len(ρup ) ≤ const · n2 ; similarly, len(ρdown ) ≤ const · n2 . Combined together, these bounds would in this case show that len(ρ) is at most quadratic in n. However, this reasoning would be insufficient for our purposes, since the number of normal arcs itself can be linear in n. This motivates more subtle upper bounds (iii) and (iv) and the fine-grained choice of parameter in (v). We show how to use Lemma 2 to obtain a quadratic upper bound on the size of the entire path in the following Subsect. 5.3; the remainder of the present subsection provides an intuitive sketch of the proof of Lemma 2. Fix configurations α and β such that cnt(α) = cnt(β) = 0 and there exists an arc from α to β. If there is a low arc from α to β, then there is nothing to prove, so assume that all the arcs from α to β are not low. Let ρ◦ be such an arc of the shortest possible length; then ρ◦ is not low. Let ρ◦ = (γ1 , t1 ) . . . (γm , tm ), t

m γm+1 = β. Since ρ◦ is shortest possible, from where α = γ1 and γm −→ Proposition 1 we infer that configurations γ1 , γ2 , . . . , γm+1 are pairwise different.

472

D. Chistikov et al.

Fig. 1. A normal decomposition of an arc

Since ρ◦ is not low, for each k = 0, 1, 2, . . . , 5n we can distinguish the first configuration γik on ρ◦ that has counter value k. Consider configurations γik for 2n ≤ k ≤ 3n. Among these configurations, some state p repeats in two configurations γik and γik , for some 2n ≤ k < k  ≤ 3n. This means that the part of ρ◦ between γik and γik corresponds to a cycle σ in the transition multigraph. This cycle has a positive effect on the counter, and hence it is contained in a positively enabled strongly connected component S ∈ S. Recall that we cannot simply pump the cycle σ in order to create ρup , because by the definition of a normal arc, the cycle that creates ρup has to be the pre-chosen cycle σS+ assigned to S. This, however, poses no real difficulty for the following reason. Since p is contained in the same strongly connected component as σS+ , we can travel through S from p to the base of σS+ , then pump it arbitrarily many times, and then go back to p. The part ρdown is defined in a symmetric manner. More precisely, the consecutive parts ρpref , ρup , ρcap , ρdown and ρsuff are defined as follows; the reader is advised to check the description against Fig. 2 while reading. First, ρpref is constructed by taking the prefix of ρ◦ up to configuration γik , and then traveling along a path σpq within S from p to q, the base state of σS+ . Then we repeat cycle σS+ a number of times, say a, thus creating ρup . The reader should think of a as of a variable, because the possibility of changing this number will be essential for the constructions to follow. Note that we chose k so that k ≥ 2n in order to make sure that during these manipulations we never hit nonpositive counter values. In a symmetrical manner we define ρsuff and ρdown . First, we find a configuration γjk¯ on ρ◦ such that it has some counter value k¯ with 2n ≤ k¯ ≤ 3n, all the configurations later on ρ◦ have smaller counter values, and its state p¯ belongs to a negatively enabled strongly connected component T . Let q¯ be the base state of σT− . Then ρsuff is the suffix of ρ◦ starting from γjk¯ , with a path σq¯p¯ appended in the front, where σq¯p¯ leads from q¯ to p¯ within T . Also, ρdown is constructed by repeating σT− a number of times, say b. We denote A = eff(σS+ ) and −B = eff(σT− ).

Shortest Paths in One-Counter Systems

473

Fig. 2. The normalization procedure applied to an arc. Parts ρpref , ρup , ρcap , ρdown , ρsuff are depicted in green, blue, grey, red, and orange, respectively, as in Fig. 1 (Color figure online)

The middle part of the path, i.e. ρcap , is constructed as follows. First, we follow a path σqp from q to p within T in order to get back to the original path. Then we repeat the cycle (σpq σqp ) exactly gcd(A, B) − 1 times, so that the total number of times paths σpq and σqp are traversed is gcd(A, B). It is necessary that the total effect on the counter value from traversing these paths is divisible by gcd(A, B), so that this effect can be compensated for by selecting a and b appropriately. Then we follow the infix of the original path up to configuration γjk¯ . Afterwards, a symmetric construction follows: we repeat the ¯ to q¯ using cycle (σp¯ ¯q σq¯p¯) exactly gcd(A, B) − 1 times, and then travel from p . It can be easily seen that one can choose numbers a and b so that the whole σp¯ ¯q construction indeed yields a path from α to β. Here we use the fact that all numbers divisible by gcd(A, B) can be represented as nonnegative combinations of A and −B, i.e., a · A − b · B for some a, b ∈ N. Finally, we modify ρcap by unpumping it “modulo gcd(A, B)” exhaustively: as long as there are two configurations that have the same state, and their counter values are congruent modulo gcd(A, B), we remove the whole subpath between these configurations. When performing such an operation, the total effect of ρcap on the counter changes by a number divisible by gcd(A, B). Hence, by adjusting once more the numbers a and b we can compensate for this change. From the description above, it should be relatively clear that the construction yields a normal path satisfying all the conditions apart from the quantitative

474

D. Chistikov et al.

ones: (iii) and (iv). For this, some arithmetic calculations are needed to ensure that we can choose a and b small enough so that (iii) and (iv) hold, while all the necessary properties of a and b are satisfied. We remark that in this sketch we have glossed over some technicalities that are used to satisfy conditions (iii) and (iv). 5.3

Length of Shortest Paths

Let α and β be such as in the statement of Theorem 1. Let ρ◦ be a path from α to β that has the minimum possible number of intermediate configurations with counter value zero. Let all these intermediate configurations with counter value zero be γ2 , . . . , γk , where γ1 = α and γk+1 = β. For i = 1, 2, . . . , k, let ρi◦ be the subpath of ρ◦ between configurations γi and γi+1 . Then ρi◦ is an arc from γi to γi+1 . By Lemma 2, there exists also an arc ρi from γi to γi+1 that is either low or is normal and admits a normal decomposition satisfying properties (i)–(vii). If ρi is low, choose ρi to be the shortest possible low arc from γi to γi+1 . If ρi is normal, let ρi = ρipref ρiup ρicap ρidown ρisuff be its normal decomposition. Our goal for the rest of the proof is to show that ρ = ρ1 . . . ρk , which is clearly a path from α to β, has length at most 14n2 . Note that ρ has the same number of configurations with counter value zero as ρ◦ . Let N ⊆ {1, 2, . . . , k} be the set of indices i for which ρi is normal, and let L = {1, 2, . . . , k} \ N be the set of indices i for which ρi is low. First we show that the sum of the lengths of low parts of ρ (more precisely, of low arcs, of ρipref and ρisuff ) is small. The following claim follows from a simple application of Proposition 1. Lemma 3 (♠). The following inequality holds:   len(ρi ) + (len(ρipref ) + len(ρisuff )) ≤ 5n2 . i∈L

i∈N

Now we will estimate the length of the rest of the path ρ. First, however, we have to prepare a toolbox of lemmas. We introduce the following notation. For i S, T ∈ S, let N(S,T ) ⊆ N be the  set of all those indices i forwhich ρ is (S, T )normal. Moreover, let N(S,·) = T  ∈S N(S,T  ) and N(·,T ) = S  ∈S N(S  ,T ) . Lemma 4 (♠). Let S, T ∈ S. Suppose i ∈ N(S,·) and j ∈ N(·,T ) for some i, j with 1 ≤ i < j ≤ k. Then there are no two configurations δi and δj appearing on ρicap and ρjcap respectively such that st(δi ) = st(δj ) and cnt(δi ) − cnt(δj ) is divisible by gcd(eff(σS+ ), −eff(σT− )). In the proof of Lemma 4 we observe that if such configurations δi and δj existed, then one could repeat σS+ on ρi and σT− on ρj more times so that the “lifted” configurations δi and δj would have the same counter value. Then we could cut the whole part of the path between them, thus reducing the number of configurations with counter value zero; this would be a contradiction with the choice of ρ◦ . The following lemma is a simple corollary of Lemma 4.

Shortest Paths in One-Counter Systems

475

Lemma 5 (♠). Let S, T ∈ S. Then |N(S,T ) | ≤ gcd(eff(σS+ ), −eff(σT− )). Total length of caps. We have now all the necessary ingredients to establish the desired upper bounds on the lengths of caps. Recall that for a strongly connected component S ∈ S we denote by nS the number of vertices in S. Lemma 6. Let S, T ∈ S, let AS = eff(σS+ ) and let BT = −eff(σT− ). Then:  len(ρicap ) ≤ AS · n; (1) i∈N(S,·)

 i∈N(·,T )

moreover ,

len(ρicap ) ≤ n · BT ; 

len(ρicap ) ≤ n2 .

(2) (3)

i∈N

 Proof. For (1), assume towards a contradiction that i∈N(S,·) len(ρicap ) > AS ·n. Then by the pigeonhole principle there exists two configurations δ and δ  on the paths ρicap for i ∈ N(S,·) which have the same state and the same counter value modulo AS . Assume w.l.o.g. that δ is earlier in the path than δ  . By property (v) of Lemma 2, configurations δ and δ  cannot appear in the same path ρicap . Indeed, otherwise the projection of the part of ρicap between δ to δ  would be a cycle with effect divisible by AS , so also by gcd(AS , −eff(σT− )), where T is the SCC for which ρi is (S, T )-normal. Therefore they have to belong to different arcs. Let δ belong to ρi and δ  belong to ρj , where j ∈ N(S,T ) for some T ∈ S. However, by Lemma 4, there are no two configurations δ and δ  on ρi and ρj , respectively, such that their states are the same and the difference in counter values is divisible by gcd(AS , −eff(σT− )). Contradiction, as δ and δ  are such configurations: the difference of its counter values is divisible by AS , so also by gcd(AS , −eff(σT− )). Thus (1) is proved, and (2) follows from a symmetric reasoning. The bound (3) follows  by summing (1) through all S ∈ S and using

the facts that eff(σS+ ) ≤ nS and S∈S nS = n. Total length of positive and negative cycles. We now show that the total sum of the lengths of ρiup and ρidown is at most 8n2 . This is the case where we need the key estimations (iii) and (iv) in Lemma 2. Lemma 7. The following inequalities hold:  len(ρiup ) ≤ 4n2 , i∈N



len(ρidown ) ≤ 4n2 .

i∈N

Proof. We show how to bound the sum of lengths of paths ρiup . For any S ∈ S, let us denote AS = eff(σS+ ) and BS = −eff(σS− ). For each i ∈ N , let Si , Ti ∈ S be such that ρi is (Si , Ti )-normal, and let Li = len(ρicap ). By Lemma 2 we know

476

D. Chistikov et al.

that eff(ρiup ) ≤ 2Li + 2 · lcm(ASi , BTi ). Since proj(ρiup ) = (σS+i )a for some integer a, we have len(ρiup ) = eff(ρiup )· Hence,



len(σS+i )

≤ eff(ρiup )·

eff(σS+i )

len(ρiup ) ≤ 2

i∈N

nSi nS ≤ (2Li +2·lcm(ASi , BTi ))· i . ASi ASi

 Li nS  lcm(AS , BT ) · nS i i i i +2 . ASi ASi

i∈N

(4)

i∈N

We will separately estimate the first and the second term. First we focus on  Li nSi i∈N AS . Let us fix some specific S ∈ S. We have i

 i∈N(S,·)

 nS nS Li nSi = · Li ≤ · AS · n = nS · n, ASi AS AS i∈N(S,·)

where the inequality follows from Lemma 6(1). Thus  Li nS  i = ASi

i∈N



S∈S i∈N(S,·)

 Li nSi ≤ nS · n = n2 . ASi

In order to estimate the second term, fix some S, T ∈ S. Note that xy y gcd(x,y)·x = gcd(x,y) for all positive integers x, y. Now we have  i∈N(S,T )

BTi · nSi = gcd(ASi , BTi )

 i∈N(S,T )

(5)

S∈S

lcm(x,y) x

=

BT · n S BT · nS = |N(S,T ) | · gcd(AS , BT ) gcd(AS , BT )

≤ gcd(AS , BT ) ·

BT · nS = BT · nS ≤ nT · nS , gcd(AS , BT )

where the first inequality follows from Lemma 5 and the second one from the fact that the effect of a path is bounded by its length. Therefore,  i∈N

  BTi · nSi BTi · nSi = gcd(ASi , BTi ) gcd(ASi , BTi ) S,T ∈S i∈N(S,T )    ≤ nT · nS = nS · nT = n2 . S,T ∈S

S∈S

(6)

T ∈S

By connecting equations (4), (5) and (6) we obtain  len(ρiup ) ≤ 2n2 + 2n2 = 4n2 . i∈N

The upper bound on the sum of lengths of paths ρidown is obtained analogously, using Lemma 6(2) instead of Lemma 6(1).

Shortest Paths in One-Counter Systems

477

Combining bounds of Lemma 3, Lemma 6(3), and Lemma 7, we conclude that len(ρ) ≤ 5n2 + n2 + 4n2 + 4n2 ≤ 14n2 , which completes the proof of Theorem 1. Acknowledgements. The authors are grateful to Christoph Haase and Aditya Kanade for discussions and comments.

References ˇ 1. Alur, R., Cern´ y, P.: Streaming transducers for algorithmic verification of singlepass list-processing programs. In: Proceedings of the 38th ACM SIGPLANSIGACT Symposium on Principles of Programming Languages, POPL 2011, Austin, TX, USA, 26–28 January 2011, pp. 599–610 (2011) 2. Atig, M.F., Bouajjani, A., Kumar, K.N., Saivasan, P.: On bounded reachability analysis of shared memory systems. In: 34th International Conference on Foundation of Software Technology and Theoretical Computer Science, FSTTCS 2014, 15–17 December 2014, New Delhi, India, pp. 611–623 (2014) 3. Barrett, C., Demri, S., Deters, M.: Witness runs for counter machines. In: Fontaine, P., Ringeissen, C., Schmidt, R.A. (eds.) FroCoS 2013. LNCS, vol. 8152, pp. 120– 150. Springer, Heidelberg (2013) 4. Bouajjani, A., Esparza, J., Maler, O.: Reachability analysis of pushdown automata: application to model-checking. In: Mazurkiewicz, A., Winkowski, J. (eds.) CONCUR 1997. LNCS, vol. 1243, pp. 135–150. Springer, Heidelberg (1997) 5. Chistikov, D.: Notes on counting with finite machines. In: 34th International Conference on Foundation of Software Technology and Theoretical Computer Science, FSTTCS 2014, 15–17 December 2014, New Delhi, India, pp. 339–350 (2014) 6. Cormen, T.H., Leiserson, C.E., Rivest, R.L., Stein, C.: Introduction to Algorithms 2nd (edn.) The MIT Press and McGraw-Hill Book Company. ISBN 0-262-03293-7 (2001) 7. Demri, S., Gascon, R.: The effects of bounding syntactic resources on Presburger LTL. J. Log. Comput. 19(6), 1541–1575 (2009) 8. Etessami, K., Wojtczak, D., Yannakakis, M.: Quasi-birth-death processes, tree-like QBDs, probabilistic 1-counter automata, and pushdown systems. Perform. Eval. 67(9), 837–857 (2010) 9. Hofman, P., Mayr, R., Totzke, P.: Decidability of weak simulation on one-counter nets. In: 28th Annual ACM/IEEE Symposium on Logic in Computer Science. LICS 2013, New Orleans, LA, USA, 25–28 June 2013, pp. 203–212 (2013) 10. Kozen, D.C.: Design and Analysis of Algorithms. Texts and Monographs in Computer Science. Springer, New York (1992). ISBN: 978-3-540-97687-5 11. Lafourcade, P., Lugiez, D., Treinen, R.: Intruder deduction for AC-like equational theories with homomorphisms. Research report LSV-04-16, Laboratoire Sp´ecification et V´erification, ENS Cachan, France, p. 69, November 2004. http:// www.lsv.ens-cachan.fr/Publis/RAPPORTS LSV/PS/rr-lsv-2004-16.rr.ps 12. Lafourcade, P., Lugiez, D., Treinen, R.: Intruder deduction for AC -like equational theories with homomorphisms. In: Giesl, J. (ed.) RTA 2005. LNCS, vol. 3467, pp. 308–322. Springer, Heidelberg (2005)

478

D. Chistikov et al.

13. Latteux, M.: Langages ` a un compteur. J. Comput. Syst. Sci. 26(1), 14–33 (1983) 14. Leroux, J., Schmitz, S.: Demystifying reachability in vector addition systems. In: 30th Annual ACM/IEEE Symposium on Logic in Computer Science. LICS 2015, 6–10 July 2015, Kyoto, Japan, pp. 56–67 (2015) 15. Podelski, A., Rybalchenko, A.: ARMC: the logical choice for software model checking with abstraction refinement. In: Hanus, M. (ed.) PADL 2007. LNCS, vol. 4354, pp. 245–259. Springer, Heidelberg (2007) 16. Thakkar, J., Kanade, A., Alur, R.: Transducer-based algorithmic verification of retransmission protocols over noisy channels. In: Beyer, D., Boreale, M. (eds.) FORTE 2013 and FMOODS 2013. LNCS, vol. 7892, pp. 209–224. Springer, Heidelberg (2013) 17. Thomas, W.: The reachability problem over infinite graphs. In: Frid, A., Morozov, A., Rybalchenko, A., Wagner, K.W. (eds.) CSR 2009. LNCS, vol. 5675, pp. 12–18. Springer, Heidelberg (2009)

The Invariance Problem for Matrix Semigroups Klaus Dr¨ ager(B) EECS, Queen Mary University of London, London, UK [email protected] Abstract. The question of whether a given subspace of Qd can be reached from a starting vector using linear transformations from a given finite set is well known to be undecidable in dimension 3 and above. We show that, in contrast, the invariance problem, i.e. the question of whether it is possible to remain inside a given subspace indefinitely using linear transformations from a given finite set, is decidable.

1

Introduction

The classic subspace reachability problem for matrix semigroups is the following. – Given: A finite set A ⊆ Md (Q) of matrices, a linear subspace U ⊆ Qd , and a vector x0 ∈ Qd . – Question: Starting from x0 , can we reach U by applying a finite sequence of matrices from A? That is, do there exist M1 , . . . , Mk ∈ A such that Mk · · · M1 x0 ∈ U ? This problem is known to be undecidable for d ≥ 3 [11]. In this paper, we consider a temporal dual, the invariance problem: – Given: A finite set A ⊆ Md (Q) of matrices, a linear subspace U ⊆ Qd , and a vector x0 ∈ Qd . – Question: Starting from x0 , can we remain in U indefinitely, using matrices from A? That is, does there exist a sequence M1 , M2 , . . . such that, for all k ∈ N, Mk ∈ A and Mk · · · M1 x0 ∈ U ? The main result of this paper is that, unlike the reachability problem, the invariance problem is decidable. The temporal duality becomes clearer when thinking about the negation of the properties: Consider the infinite-state transition system S = (Qd , x0 , A) with transitions Qd → Qd defined by the given set A of matrices, and the formula ϕU ≡ r1 · x = . . . = rk · x = 0 defining the subspace U in terms of a basis (r1 , . . . , rk ) of its orthogonal complement, then a solution to the reachability problem is a counterexample to the LTL assertion S  ¬ϕU (stating that S will always be outside U ), while a solution to the invariance problem is a counterexample to the LTL assertion S  ♦¬ϕU (stating that S will eventually be outside U ). Note the existential character of the problem: we are asking about satisfiability of the invariant. The universal version (the question of validity) can easily be shown to be also decidable; see Sect. 4.3. c Springer-Verlag Berlin Heidelberg 2016  B. Jacobs and C. L¨ oding (Eds.): FOSSACS 2016, LNCS 9634, pp. 479–492, 2016. DOI: 10.1007/978-3-662-49630-5 28

480

K. Dr¨ ager

As a technical aside, when we talk about a matrix semigroup in this paper, we mean a sub-semigroup G of Mn (Q) equipped with a particular generating set A. For the reachability problem, this detail is irrelevant (all that is required is the existence of some M ∈ G with M x0 ∈ U ), but it matters for the invariance problem. Consider the case x0 = (1, 1, 0)T , U = {(x, y, 0)T | x, y ∈ Q}, and A1 = {M1 , M3 }, A2 = {M2 , M3 }, where ⎛ ⎞ ⎛ ⎞ ⎛ ⎞ 010 001 100 M1 = ⎝1 0 0⎠ , M2 = ⎝0 1 0⎠ , M3 = ⎝0 0 1⎠ . 001 100 010 Both A1 and A2 generate the natural permutation representation of the symmetric group S3 , but the invariance problem for x0 , U, A1 has a solution M1ω ∈ Aω 1, while there is no solution for x0 , U, A2 . 1.1

Attacking the Invariance Problem

There is an intuitively obvious approach to the invariance problem, which proceeds as follows: – Expand the tree of words w ∈ A∗ . – Abort any branch M1 . . . Mk with Mk · · · M1 x0 ∈ / U ; if there are no more branches left, then there is no solution. – If for some branch M1 . . . Mk we have Mk · · · M1 x0 = 0, we have a solution, since 0 ∈ U is fixed by any M ∈ A (i.e. M1 . . . Mk w will do for any w ∈ Aω ). However, this method is not complete. The result which allows us to fix it is a pumping lemma for solution prefixes: there is a bound N depending only on |A| and dim U such that any word w ∈ Aω has a prefix uv  w with – |uv| ≤ N , and – (u, v) is dominating in a sense defined below, which in particular implies that v is nonempty and if w is a solution to the invariance problem, then so is uv ω . Once we have this, the decision problem reduces to checking finitely many pairs (u, v), as in the algorithm in Fig. 1. Example 1. Let x0 = A = {P, Q, R} with ⎛ 1 2 0 ⎜0 0 1 P =⎜ ⎝2 1 0 0 0 2

(1, 0, 0, 0)T ∈ Q4 , U = {(x, y, z, 0)T | x, y, z ∈ Q}, and ⎞ ⎛ ⎞ ⎛ 0 0 −1 0 2 0 ⎜−1 0 2 0 ⎟ ⎜−1 2⎟ ⎟,Q = ⎜ ⎟ ⎜ ⎝ 2 0 −1 0 ⎠ , R = ⎝ 0 0⎠ 1 0 2 0 −1 2

⎞ 0 −1 2 0 2 0⎟ ⎟. 1 0 0⎠ 0 0 −1

Figure 2 shows the beginning of a breadth-first exploration for this example; one solution to the invariance problem is the periodic sequence P QR(RP P QP Q)ω , corresponding to the dominating prefix (P QR, RP P QP Q).

The Invariance Problem for Matrix Semigroups

481

Fig. 1. Finding a trace remaining inside U , starting from x0 . Exploration follows a breadth-first strategy in order to find a minimal-length solution. The details of the notions of pumpability and dominating prefix, and proof that such a prefix exists for any sufficiently long branch, are given in Sects. 2 and 3.

Fig. 2. The first few layers of the running example

482

2

K. Dr¨ ager

Preliminaries

2.1

Setting

Throughout this paper, we work within Qd for some d ∈ N. We assume the linear subspace U ⊆ Qd , starting vector x0 ∈ U , and finite non-empty set of matrices A ⊆ Md (Q) to be arbitrary but fixed, subject to the non-triviality condition M U  U for all M ∈ A. If there is some M violating this condition, then M ω is a trivial solution to the invariance problem. We use A∗ and Aω for the sets of finite and infinite words over A, respectively, and  for the empty word. For x ∈ Qn , S ⊆ Qn , and w = M1 . . . Mk ∈ A∗ , w(x) := Mk · · · M1 x and w(S) := {w(x) | x ∈ S}. A word p ∈ A∗ is a prefix of w = M1 . . . Mk ∈ A∗ (resp. w = M1 M2 . . . ∈ Aω ), denoted by p  w, iff p = M1 . . . Mj for some j ≤ k (resp. some j ∈ N). Definition 1. Let w = M1 . . . Mk ∈ A∗ . Associated with w are the following subspaces of Qn : – the source space S(w) = {x ∈ Qn | p(x) ∈ U for each prefix p  w}, and – the target space T (w) = w(S(w)). The following elementary properties of S( ), T ( ) are used throughout the proof of the main result: Lemma 1. For all u, v, w ∈ A∗ , (i) S(uv) = {x ∈ S(u) | u(x) ∈ S(v)} ⊆ S(u), with equality iff T (u) ⊆ S(v), (ii) T (uv) = v(T (u) ∩ S(v)) ⊆ T (v), with equality if T (u) ⊇ S(v), (iii) dim T (uvw) ≤ dim T (v). Proof (i) Since any prefix of uv is either a prefix of u or a word of the form up, where p is a prefix of v, we have S(uv) = {x ∈ Qn | p(x) ∈ U for all p  uv} = {x ∈ Qn | p(x) ∈ U for all p  u and p(u(x)) ∈ U for all p  v} = {x ∈ Qn | x ∈ S(u) and u(x) ∈ S(v)}, which is obviously a subspace of S(u). Equality holds iff for all x ∈ S(u), u(x) is in S(v), iff T (u) ⊆ S(v). (ii) Using uv(x) = v(u(x)), we get T (uv) = v(u(S(uv))) = v(u({x ∈ S(u) | u(x) ∈ S(v)})) = v({u(x) | x ∈ S(u) and u(x) ∈ S(v)}) = v(T (u) ∩ S(v)) ⊆ v(S(v)) = T (v). If T (u) ⊇ S(v), the inclusion in the last line holds with equality.

The Invariance Problem for Matrix Semigroups

483

(iii) By (ii), T (uvw) ⊆ T (vw), and we have dim T (uvw) ≤ dim T (vw) = dim w(T (v) ∩ S(w)) ≤ dim w(T (v)) ≤ dim T (v), where in the last step we use that dim f (V ) ≤ dim V for any linear map f and vector space V .

If w = M1 M2 . . . ∈ Aω is an infinite word, then due to Lemma 1(i), the source spaces of its prefixes form a descending chain S() ⊇ S(M1 ) ⊇ . . .; since the lattice of subspaces satisfies the descending chain condition, this chain has a limit, S(w), the space of all vectors whose w-orbit remains in U . The invariance problem obviously amounts to checking the existence of an infinite word w with x0 ∈ S(w). Definition 2. A word q ∈ A∗ is pumpable if q =  and T (q) ⊆ S(q). A dominating prefix of a (finite or infinite) word w is a pair (p, q) such that pq is a proper prefix of w and q is pumpable. Any word which has a dominating prefix can then be disregarded due to the following lemma. Lemma 2. Let p, q ∈ A∗ . If q is pumpable, then (i) S(q k ) = S(q) for all k ≥ 1, (ii) S(pq k ) = S(pq) for all k ≥ 1, and (iii) if pq is a prefix of a word w ∈ Aω , and w is a solution to the invariance problem for x0 , U, A, then so is pq ω . Proof (i) We use induction on k. The assertions holds trivially for k = 1; for the step k → k + 1, from T (q) ⊆ S(q) = S(q k ) we get, using Lemma 1(i), that S(q) = S(qq k ) = S(q k+1 ). (ii) From (i) and Lemma 1(i) we have S(pq) = {x ∈ S(p) | p(x) ∈ S(q)} = {x ∈ S(p) | p(x) ∈ S(q k )} = S(pq k ). (iii) S(pq ω ) is the limit of (S(pq k ))k≥1 , which by (ii) equals the constant sequence (S(pq))k≥1 , and therefore S(pq ω ) = S(pq). Since pq is a prefix of w, and x0 ∈ S(w), we get

x0 ∈ S(w) ⊆ S(pq) = S(pq ω ). Our goal now is to show that there is a bound N such that any word w with |w| > N has a dominating prefix (p, q) with |pq| ≤ N . Definition 3. A word w ∈ A∗ is – essential if it has no dominating prefix, – S-minimal of dimension k if S(w)  S(p) for each proper prefix p  w and dim S(w) = k,

484

K. Dr¨ ager

– T -minimal of dimension k if k = dim T (w) < dim T (p) for each proper prefix p  w, – S-essential (resp. T -essential) of dimension k if it is both essential and Sminimal (resp. T -minimal) of dimension k. We will omit the dimension for the last three properties when it is irrelevant. Example. Continuing from the partial exploration shown in Fig. 2, Table 1 contains the source and target spaces for some selected words occurring in the tree, and whether or not they are S-minimal. All the words in the figure are essential. Table 1. Some source and target spaces occurring in the running example w

S(w)

T (w)



{(x, y, z, 0)T }

{(x, y, z, 0)T }

T

S-minimal yes, dimension 3 T

P

{(x, y, 0, 0) }

{(x + 2y, 0, 2x + y, 0) }

yes, dimension 2

Q

{(x, 0, z, 0)T }

{(0, 2z − x, 2x − z, 0)T }

yes, dimension 2

T

T

R

{(0, y, z, 0) }

{(−z, 2z, y, 0) }

yes, dimension 2

PQ

{(x, y, 0, 0)T }

{(0, 3x, 3y, 0)T }

no

QR

{(x, 0, z, 0)T }

{(z − 2x, 4x − 2z, 2z − x, 0)T }

no

T

T

P QP

{(x, 0, 0, 0) }

{(6x, 0, 3x, 0) }

yes, dimension 1

P QR

{(x, y, 0, 0)T }

{(−3y, 6y, 3x, 0)T }

no

P QRQ

T

{(x, 0, 0, 0) }

T

{(0, 6y, −3x, 0) }

yes, dimension 1

The plan now is to prove that there are only finitely many essential words, using the following factorization. Definition 4. The S-factorization F (w) of a word w ∈ A∗ is a finite sequence of nonempty finite words defined as follows. If w = , then F (w) = (). For w = , let p  w be the shortest prefix of w for which S(p) = S(w), and (q1 , . . . , qk ) = F (q), where q is the suffix with w = pq; then F (w) = (p, q1 , . . . , qk ). Note that since we require M U  U for all M ∈ A, we have U = S()  S(w) for all w = , which ensures that p = . For infinite w ∈ Aω , we get an infinite version of this factorization by corecursively defining F (w) = (p1 , p2 , . . .), where p1 is the shortest prefix of w with S(p1 ) = S(w), and (p2 , . . .) = F (q) for the infinite suffix q with w = pq. Note that actually computing the infinite factorization (even incrementally) would generally not be feasible (since a decrease in dimension S(uM )  S(u) could occur after an arbitrarily long prefix u, we could not determine p1 based on any finite prefix u, unless S(u) happens to be {0}), but we will just need its existence and some of its properties. However, in case w is periodic, we can compute F (w):

The Invariance Problem for Matrix Semigroups

485

Example 2. Consider the word w = uv ω with u = P QR, v = RP P QP Q from Example 1. A quick calculation gives that S(v) = T (v) = S(v ω ) = {(0, 0, z, 0)T | z ∈ Q}. Since u(x0 ) ∈ S(v), we get that x0 ∈ S(p), and therefore dim S(p) ≥ 1, for any prefix p  w. On the other hand, since P, Q, R are invertible, we have that dim S(pv) = dim T (pv) ≤ dim T (v) = 1 for any word p, so that the dimension of each factor in F (w) must be 1. Since w is periodic, the same will be true for the factorization. In fact, we get F (P QR(RP P QP Q)ω ) = (P QR, RP, P Q, P QR, P P Q, P QR, . . .)

3

Deciding Satisfiability

We now show that there are only finitely many essential words, and derive upper bounds on their number and length, from which the main result follows. 3.1

Finiteness of the Set of Essential Words

Lemma 3. If w ∈ A∗ is S-minimal of dimension k, then it is also T -minimal of some dimension j ≤ k. As a direct corollary, if w is S-essential of dimension k, then it is also T -essential of some dimension j ≤ k. Proof. That dim T (w) ≤ dim S(w) follows directly from T (w) = w(S(w)). If w = , then it is trivially T -minimal. Otherwise, w = vM for some v ∈ A∗ and M ∈ A. By Lemma 1(iii), it suffices to show dim T (w) < dim T (v). Since w is S-minimal, S(w) = S(vM )  S(v), and thus T (v)  S(M ) by Lemma 1(i). Therefore dim T (vM ) = dim M (T (v) ∩ S(M )) ≤ dim (T (v) ∩ S(M ))

by Lemma 1(ii)

< dim T (v), i.e. w is T -minimal. The corollary follows by just adding non-existence of dominating prefixes of w to both the assumption and conclusion.

Theorem 1. Let w = vM be a non-empty essential word, where v ∈ A∗ and M ∈ A. Let the S-factorization of v be F (v) = (p1 , . . . , pm ). Then (i) each pi is T -essential, (ii) if w is T -minimal of dimension k, then each pi is T -minimal of some dimension di > k, and (iii) pi = pj for i = j.

486

K. Dr¨ ager

Proof (i) By construction of the S-factorization, each pi is S-minimal. By Lemma 3, it is also T -minimal. Suppose pi is not essential, then it has a dominating prefix (r, s). This implies that (p1 . . . pi−1 r, s) is a dominating prefix for the essential word w, contradiction. (ii) We already have that each pi is S-minimal and thus T -minimal by Lemma 3; it remains to show dim T (pi ) > dim T (w). This follows because pi is a factor of v, so dim T (pi ) ≥ dim T (v) by Lemma 1(iii), and dim T (v) > dim T (w) due to T -minimality of w. (iii) Suppose pi = pj for some i < j. By the definition of F (v) we then have S(pi . . . pm ) = S(pi ) = S(pj ) = S(pj . . . pm ) ; furthermore, since the spaces (S(pi . . . pk )) form a descending chain, we have in fact S(pi ) = S(pi . . . pk ) for all i ≤ k ≤ m, and in particular S(pi ) = S(pi . . . pj−1 ) = S(pi . . . pm ). This implies T (pi . . . pj−1 ) ⊆ S(pj . . . pm )

by Lemma 1(i)

= S(pj ) = S(pi ) = S(pi . . . pj−1 ). Therefore (p1 . . . pi−1 , pi . . . pj−1 ) is a dominating prefix for w, contradiction.

This theorem allows us to prove finiteness of the set of essential words by induction on the codimension c(w) := dim U − dim T (w). It also enables us to derive upper bounds on their number and length, for which we need the following definition. Definition 5 – The arrangement function a : N → N is given by a(n) =

n  n! i=0

i!

;

it is the number of sequences from a set of n elements with no repeated element. Note that a(n)/n! converges to Euler’s number e from below, in particular a(n) ≤ 3n! for all n. 1 for i = 0 – The numbers Ni for i ∈ N are defined by Ni = |A| · a(Ni−1 ) otherwise. 0 for i = 0 – The numbers Li for i ∈ N are defined by Li = Ni−1 Li−1 + 1 otherwise. Theorem 2. We have the following bounds on the numbers and lengths of essential words w, based on their codimension c(w) = dim U − dim T (w).

The Invariance Problem for Matrix Semigroups

487

(i) There are at most Ni T -essential words w of codimension c(w) ≤ i. (ii) A T -essential word of codimension i has length ≤ Li . (iii) There at most N1+dim U essential words, and none of them is longer than L1+dim U . Proof (i) We proceed by induction on i. For i = 0, the empty word  is trivially T essential. Since T () = U and   w for all w ∈ A∗ , it is the only T -essential word of dimension dim U , i.e. codimension 0. For i → i + 1, by the induction hypothesis there are at most Ni T -essential words of codimension c(w) ≤ i. By Theorem 1, any T -essential word w of codimension i + 1 has a factorization w = p1 . . . pn M in which the pj are pairwise distinct and T -essential of codimension ≤ i, and M ∈ A. The number of such factorizations is at most |A|·a(Ni ) = Ni+1 , and they include the ones for words of codimension ≤ i, giving the upper bound Ni+1 for the number of T -essential words w of codimension c(w) ≤ i + 1. (ii) As in (i), we argue inductively using Theorem 1. For i = 0, the only T essential word  of codimension 0 has length 0 = L0 . For i → i+1, the decomposition in Theorem 1 consists of at most Ni words, each of length at most Li , plus the final letter, giving a total length of at most Ni Li + 1 = Li+1 . (iii) As in (i) and (ii), we get from Theorem 1 that any non-empty essential word has a factorization into (at most Ndim U ) pairwise T -essential words of length at most Ldim U and a single trailing M ∈ A. The number of such factorizations is at most N1+dim U , and their length is bounded by L1+dim U , by the same argument as before.

3.2

Decidability of the Invariance Problem

The main result now follows immediately from the bounds established in the previous section. Theorem 3. Algorithm 1 terminates. It returns FAIL if and only if there is no solution to the invariance problem. Proof. Let w = M1 M2 . . . ∈ Aω be any infinite word. By Theorem 2, w has an essential prefix m(w) = M1 . . . Mk of maximal length |m(w)| ≤ L1+dim U (note that the empty word is always essential, so m(w) exists). Then M1 . . . Mk+1 has a dominating prefix (p, q). We must have that pq = m(w) and there is no shorter dominating prefix, since otherwise m(w) would also fail to be essential. Due to the properties of dominating prefixes, x0 ∈ S(m(w)) if and only if pq ω is a solution to the invariance problem. Therefore we have two cases. If there is no solution to the invariance problem, then for any branch w, x0 cannot be in S(m(w)), i.e. the branch will be discarded before reaching depth onig’s lemma, only finitely many words are explored, |m(w)| ≤ L1+dim U . By K¨ and the algorithm returns FAIL.

488

K. Dr¨ ager

If there is a solution w, then x0 ∈ S(p) for every prefix p  w, and the algorithm keep exploring until it reaches depth |m(w)| for one such w, at which point the dominating prefix is discovered and returned.

4 4.1

Further Remarks and Variations Computing All Possible Initial Vectors

The algorithm we gave can be easily adapted to solve the following, more general problem: – Given: A finite set A ⊆ Md (Q) of matrices, and a subspace U ⊆ Qd . – Question: From which starting vectors x0 can we remain in U indefinitely, using matrices from A? That is, for which x0 ∈ Qd does there exist a sequence M1 , M2 , . . . such that, for all k ∈ N, Mk ∈ A and Mk · · · · · M1 x0 ∈ U ? Essentially, all that has to be changed is to remove the special treatment of x0 and collect all the pairs which would have been returned into a set P , as in Fig. 3.

Fig. 3. Finding all starting vectors from which the invariance problem has a solution.

4.2

Locations

The transition system (Qd , x0 , A) can be extended using a finite set L of control locations, giving (L × Qd , (l0 , x0 ), T ) for a finite set T ⊆ L × Mn (Q) × L. The main properties of words w ∈ A∗ can still be used as before, the main difference being that a pumpable word q is now additionally required to label a cycle in the location graph. The algorithm then proceeds as before, except that nodes are labeled with states (l, x) ∈ L × Qd , and only successors using matrices which are available in l are considered.

The Invariance Problem for Matrix Semigroups

4.3

489

The Universal Version

The question whether the invariant given by U is valid, i.e. whether w(x0 ) ∈ U for all w ∈ A∗ , is also decidable, and is in fact easier than the problem we have dealt with in the previous sections. Again, the basic idea is to expand the tree of words w ∈ A∗ in a breadth-first order. While doing so, we can (i) as soon as we encounter a counterexample w(x0 ) ∈ / U , return w; (ii) cut any branch v if v(x0 ) is a linear combination of vectors we have previously explored. The reason for (ii) is that, if y := v(x0 ) = λ1 x1 +· · ·+λn xn , then due to linearity / U for any w ∈ A∗ , w(y) = λ1 w(x1 ) + · · · + λn w(xn ); in particular, if w(y) ∈ for some w ∈ A∗ , then there is some i for which w(xi ) is also not in U . Since we use breadth-first search, xi = ui (x0 ) for some word ui which is smaller than v in the length-lexicographic order. Thus for any counterexample which we lose by discarding y, there is a length-lexicographically shorter one. In particular, if there are any counterexamples, then the length-lexicographically minimal one among them cannot be lost and will be found by the search.

Fig. 4. Checking validity of the invariant U , starting from x0 .

5 5.1

Summary and Future Work Summary

We presented a solution to the invariance problem for matrix semigroups, i.e. the question of an infinite sequence of matrices satisfying a given linear invariant.

490

K. Dr¨ ager

We gave an algorithm to find a solution if one exists, and proved its termination. The latter relied on the analysis of various properties of words in matrix semigroups and their interaction. In particular, we showed that for any finite set A of matrices and subspace U , there is a bound N such that any word of length > N has a dominating prefix, reducing the problem to a finite one (Fig. 4). 5.2

Related Work

Matrix semigroups are a rich source of problems [3, 11], including many surprisingly complicated decision problems. Among these are the scalar reachability problem (undecidable in dimension 3 and above [3]), which is the problem of reaching a subspace of codimension 1 from a starting vector using matrices from a given semigroup, and the vector reachability problem, in which the target is a single vector. A special case of the latter, for a single generating matrix, is the orbit problem which was shown to be decidable by Kannan and Lipton [5]. A closely related problem to the universal version of the invariance problem (Sect. 4.3) is the question of boundedness, which is undecidable [1]. One source of the complexity of such problems is that, in contrast to similar models like vector addition systems [4,6, 9], the behaviour of matrix semigroups has nonlinear aspects; for example, the simple 3-dimensional matrix ⎛ inherently ⎞ 121 ⎝0 1 1⎠ has the orbit {(n2 , n, 1)T | n ∈ N}. This ability to reflect polynomial 001 relationships connects problems like scalar reachability to known undecidable ones, like solvability of diophantine equations [10]. This complexity extends to various related models. Polynomial Register Machines [2] generalize vector addition systems with polynomial update functions (with integer coefficients); while this leads to undecidability in most cases, in dimension 1 reachability turns out to be decidable (in fact PSPACE-complete). Iteration of piecewise affine maps [7, 8], similarly to matrix semigroups, involves a choice between affine-linear transformations, but this choice is deterministic based on the current variable values; the (vector) reachability problem is undecidable in general, but decidability is still open in dimension 1. 5.3

Future Work

There are a number of interesting ways to extend these results. Among them are: Complexity Bounds. From the proof of decidability we get the upper bound L∞ ≤ 1 + Ndim U (1 + Ndim U −1 (. . . (1 + N0 ) . . .) for the exploration depth, where N0 = 1 and Ni+1 ≤ 3·|A|·Ni !, for a total of dim U nested factorials. This implies a complexity upper bound of O(2L∞ ) for the algorithm. It would be interesting to see how much this can be improved, and what lower bounds can be found.

The Invariance Problem for Matrix Semigroups

491

Polyhedral Invariants. The invariant in this paper was given as a linear subspace U , corresponding to a conjunction of linear equations. Generalizing this to linear inequalities leads to the question of whether we can find a sequence of matrices from A which allows us to remain inside a given polyhedron. This adds significant complications; in particular, it is not the case that a descending sequence of polyhedra P0 ⊇ P1 ⊇ . . . stabilizes after finitely many steps, so notions like the S-factorization cannot simply be translated to this setting. It is not clear at all whether this more general question is decidable. Guards. The control structure of the transition system (Qd , x0 , A) is relatively simple in that any transition is enabled at any time. Adding control location already changes this, but the values of the vector x in a state (l, x) still have no influence on the control flow. This changes with the introduction of application conditions or guards: linear equations or inequalities which have to be satisfied before the associated transition can be taken. Note that equation guards can be translated into extra dimensions in such a way that guard violations translate into (extended) invariance violations, so that they don’t increase expressivity; inequalities on the other hand make for an interesting addition. Games. Since both the existential and (as seen in Sect. 4.3) the universal version of the invariance problem are decidable, it is natural to ask what can be done about an alternating version. This would amount to considering games in which two players , ♦ take turns applying matrices from given sets A , A♦ to the current vector x, with the goal of preserving and violating the invariant U , respectively. Acknowledgements. The EP/K032011/1.

author

is

part

supported

by

EPSRC

project

References 1. Blondel, V.D., Tsitsiklis, J.N.: The boundedness of all products of a pair of matrices is undecidable. Syst. Control Lett. 41(2), 135–140 (2000) 2. Finkel, A., G¨ oller, S., Haase, C.: Reachability in register machines with polynomial updates. In: Chatterjee, K., Sgall, J. (eds.) MFCS 2013. LNCS, vol. 8087, pp. 409– 420. Springer, Heidelberg (2013) 3. Halava, V., Harju, T., Hirvensalo, M.: Undecidability bounds for integer matrices using claus instances. Technical report 766 (2006) 4. Hopcroft, J., Pansiot, J.-J.: On the reachability problem for 5-dimensional vector addition systems. Theor. Comput. Sci. 8(2), 135–159 (1979) 5. Kannan, R., Lipton, R.J.: Polynomial-time algorithm for the orbit problem. J. ACM 33(4), 808–821 (1986) 6. Karp, R.M., Miller, R.E.: Parallel program schemata. J. Comput. Syst. Sci. 3(2), 147–195 (1969) 7. Kurganskyy, O., Potapov, I.: Reachability problems for pams. CoRR, abs/1510.04121 (2015)

492

K. Dr¨ ager

8. Kurganskyy, O., Potapov, I., Sancho-Caparrini, F.: Reachability problems in lowdimensional iterative maps. Int. J. Found. Comput. Sci. 19(04), 935–951 (2008) 9. Leroux, J.: Vector addition systems reachability problem (a simpler solution). In: Voronkov, A. (ed.)Turing-100. The Alan Turing Centenary. EasyChair Proceedings in Computing, vol. 10, pp. 214–228. EasyChair (2012) 10. Matiyasevich, Y.: Hilbert’s Tenth Problem. With a foreword by Martin Davis. MIT Press, Cambridge (1993) 11. Potapov, I.: From post systems to the reachability problems for matrix semigroups and multicounter automata. In: Calude, C.S., Calude, E., Dinneen, M.J. (eds.) DLT 2004. LNCS, vol. 3340, pp. 345–356. Springer, Heidelberg (2004)

Order-Sorted Rewriting and Congruence Closure Jos´e Meseguer(B) Department of Computer Science, University of Illinois at Urbana-Champaign, Urbana, USA [email protected]

Abstract. Order-sorted type systems supporting inheritance hierarchies and subtype polymorphism are used in theorem proving, AI, and declarative programming. The satisfiability problems for the theories of: (i) order-sorted uninterpreted function symbols, and (ii) of such symbols modulo a subset Δ of associative-commutative ones are reduced to the unsorted versions of such problems at no extra computational cost. New results on order-sorted rewriting are needed to achieve this reduction. Keywords: Order-sorted rewriting · Congruence closure · Satisfiability

1

Introduction

For greater expressiveness and efficiency, type systems supporting inheritance hierarchies and subtype polymorphism are used in many areas such as resolution theorem proving, e.g., [26, 32], declarative logic and rule-based languages, e.g., [4,9,10, 29], and artificial intelligence, e.g., [8, 29]. Order-sorted (OS) equational logic, e.g., [15,21], is a logical framework supporting inheritance hierarchies and subtype polymorphism widely used for these purposes. Therefore, the development of decision procedures for OS theories is of interest in all these areas. I focus here on decision procedures for the OS theory of uninterpreted function symbols, which in an unsorted setting is decided by congruence closure algorithms [7, 24,27]. However, for greater expressiveness one can allow some of the function symbols, say in a subsignature Δ ⊆ Σ, to be interpreted by some axioms BΔ . For example, for an unsorted subsignature Δ ⊆ Σ of binary function symbols, congruence closure algorithms modulo the axioms ACΔ , asserting the associativity and commutativity of all symbols in Δ have been given in [2, 19, 22]. Therefore, I also study satisfiability in the OS theory (Σ, ACΔ ) of uninterpreted function symbols Σ modulo ACΔ . The most obvious approach would be to develop an order-sorted congruence closure algorithm along the lines of [11] and then extended it to the modulo AC case. However, the main, somewhat surprising message of this paper is that such OS congruence closure algorithms are not needed at all : the already existing and efficient unsorted congruence closure algorithms in [7, 24, 27] and congruence closure modulo ACΔ in [2, 19,22] and tools supporting them can be reused without change and at no extra cost to solve the corresponding OS satisfiability problems. c Springer-Verlag Berlin Heidelberg 2016  B. Jacobs and C. L¨ oding (Eds.): FOSSACS 2016, LNCS 9634, pp. 493–509, 2016. DOI: 10.1007/978-3-662-49630-5 29

494

J. Meseguer

A Simple Example. Consider the following order-sorted signature Σ

with sorts A, B, C, subsorts A, C < B, f subsort-polymorphic with typings f : A → A and f : C → C, and a binary + with typing + : A A → A. Its socalled theory of uninterpreted function symbols is just the order-sorted equational theory (Σ, ∅) with empty set of equations, whose class of models, OSAlgΣ , is that of all order-sorted Σ-algebras detailed in Sect. 2. Is the formula ()

a = b ∧ b = c ∧ f (f (a)) = f (a) ∧ a + f (f (a)) = f (a) + a

(Σ, ∅)-satisfiable? The standard way to answer this question if Σ were unsorted would be to: (1) compute the congruence closure of the first three equations; and (2) test the last inequality using such a congruence closure. Since, as pointed out in [2, 12,16], unsorted congruence closure algorithms are ground Knuth-Bendix completion algorithms [18], an obvious way to try to answer this question would be to try to complete the first three equations into an equivalent set of confluent and terminating rewrite rules. But this runs into serious trouble. An ordersorted Knuth-Bendix completion algorithm such as [13] will orient a = b and b = c as b → a and b → c because rules must be sort-decreasing, i.e., rewrite to a term of equal or lower sort. This then generates the critical pair a = c, which is unorientable, so completion fails. Notice also that replacement of equals by equals does not hold in an order-sorted setting: from a = b we cannot derive f (a) = f (b), because f (b) doesn’t type. These difficulties were clearly felt by the authors of [11], the only order-sorted congruence closure algorithm I am aware of, which is quite complex and is not a Knuth-Bendix completion. They say: An approach using rewriting [. . .] fails due to the well-known problem that rewriting with order-sorted rewrite rules may create ill-typed terms. Let us now widen the problem into one of satisfiability modulo AC by making the + symbol associative-commutative. That is, we consider the axioms AC + = {x+ y = y + x, (x + y) + z = x + (y + z)}, with x, y, z of sort A, and ask: is the formula () (Σ, AC + )-satisfiable? For this case, I am not aware of any order-sorted ACcongruence closure algorithm, but unsorted, ground-AC-completion-based ones exist [2, 19, 22]. The trouble, again, is that order-sorted AC-completion as in [13] fails miserably in the same way (a = c cannot be oriented).

Order-Sorted Rewriting and Congruence Closure

495

Wouldn’t it be nice if we could completely ignore all sort information in the above two OS satisfiability problems and solve them as unsorted problems using standard (and efficient!) congruence closure [7, 24, 27] and congruence closure modulo AC [2,19, 22] algorithms? If this reduction method were sound, we could easily settle the (Σ, ∅)- and (Σ, AC + )-satisfiability of (): the confluent and terminating rules R = {a → b, c → b, f (f (b)) → f (b)} play the role of a “congruence closure” for the first three equations, and also of an AC+ -congruence closure. Since the disequality a + f (f (a)) = f (a) + a reduces to b + f (b) = f (b) + b, the formula () is (Σ, ∅)-satisfiable. However, since b + f (b) =AC+ f (b) + b, () is (Σ, AC + )-unsatisfiable. But is this reduction to unsorted satisfiability sound ? Initial Algebra Semantics of Uninterpreted Satisfiability. Ignoring the sort information of an OS signature Σ is captured by a signature map u : Σ  (f : s1 . . . sn → s) → (f : U . n. . U → U ) ∈ Σ u , where U is the single “universe” sort in the unsorted signature Σ u . As further detailed at the end of Sect. 2, u induces a reduct map of algebras in the opposite direction, |u : AlgΣ u  A → A|u ∈ OSAlgΣ , making each unsorted algebra A into and order-sorted one A|u , and such that for a set of ground OS Σ-equations E we have the equivalence: A|u |= E ⇔ A |= E. In particular, the E-initial unsorted Σ u -algebra TΣ u /E is mapped to the OS Σ-algebra TΣ u /E |u and, since TΣ u /E |u |= E, there is a unique OS homomorphism h : TΣ/E → TΣ u /E |u from the E-initial OS Σ-algebra TΣ/E . But the proof ofTheorem   5 shows that, for equations E and disequations D, the conjunction E ∧ D is satisfiable iff TΣ(C)/E |= E ∧ D, where the variables C of E ∪ D are seen as  fresh new  constants added to Σ to get a supersignature Σ(C) ⊇ Σ, so that E ∧ D becomes a ground formula. This gives us, in model-theoretic terms, the key to verify the soundness of the hoped-for reduction of the satisfiability for the theory of OS uninterpreted function symbols to that of the unsorted theory of uninterpreted function symbols: this reduction method will be sound if and only if the OS homomorphism h : TΣ(C)/E → TΣ(C)u /E |u is injective. In proof-theoretic terms this injectivity will hold if and only if for all ground Σ-equation u = v we have the equivalence: (Σ, E) u = v ⇔ (Σ u , E) u = v. The (⇒) direction is obvious, but the (⇐) direction is a non-trivial new result that follows from several conservativity theorems that I prove in Sects. 3.2 and 4.1 by factoring the signature  → Σ u of increasingly map u : Σ → Σ u through a sequence Σ → Σ  → Σ simpler order-sorted, many-sorted and finally unsorted signatures and relating equational and rewriting deductions at all these levels. The Plot Thickens. The soundness of the hoped-for reduction to the unsorted case for satisfiability modulo ACΔ is a thornier issue. As before, the reduction will be sound if and only if for ground Σ-equations E the unique Σhomomorphism h : TΣ/E∪ACΔ → TΣ u /E∪ACΔu |u from the initial E ∪ ACΔ algebra TΣ/E∪ACΔ is injective. But some of the conservativity theorems along  → Σ u needed to make h the above sequence of signature maps Σ → Σ  → Σ injective actually break down in the ACΔ case. The problem has to do with the translation of the equations ACΔ along these signature maps. At the unsorted

496

J. Meseguer

level of Σ u the translated equations ACΔu , are more general and therefore identify more terms than the original OS equations ACΔ . Consider a simple example: the equation a + b = b + a does not type in our example signature Σ, but it types in the supersignature Σ  ⊇ Σ, which for our running example is depicted in Sect. 3.1. The AC equations ACΔ in our example are just associativity and commutativity of + : A A → A and therefore apply only to terms of sort A. Instead, the AC equations ACΔu are unsorted, and apply to all terms. This means that a + b =ACΔu b + a, but since b does not have sort A, we have a + b =ACΔ b + a. It also means that the homomorphism h : TΣ  /E∪ACΔ → TΣ u /E∪ACΔu |u in general is not injective. However, all hope is not lost. As a direct consequence of Corollary 2 in Sect. 3.2, there is an isomorphism α : TΣ/E∪ACΔ ∼ = TΣ  /E∪ACΔ |Σ to the Σ-reduct of TΣ  /E∪ACΔ and this shows that the homomorphism h : TΣ/E∪ACΔ → TΣ u /E∪ACΔu |u that we need to prove injective for the reduction to be sound is up to isomorphism a restriction of h to TΣ/E∪ACΔ , which could be injective even if h is not. Lemma 3 in Sect. 4.1 and the highly non-trivial Theorem 8 in Sect. 5 save the day: it follows from them that h is indeed injective and the reduction is also sound for the AC case. To the best of my knowledge the results on reducing order-sorted to unsorted satisfiability and on order-sorted rewriting and equality are new. The paper is organized as follows. After some preliminaries in Sect. 2, the new results on order-rewriting and equality are given in Sect. 3. The reductions of satisfiability in the theory of OS uninterpreted function symbols (resp. OS uninterpreted function symbols modulo AC) to satisfiability in their respective unsorted theories is given in Sect. 4 (resp. Sect. 5). Related work and conclusions are discussed in Sect. 6. Due to space limitations no proofs are given; they can be found in the Technical Report [20].

2

Preliminaries on Order-Sorted Algebra

The following material is adapted from [21], which generalizes [15]. It summarizes the basic notions of order-sorted algebra needed in the rest of the paper. Definition 1. A many-sorted signature is a pair Σ = (S, Σ), with S a set of sorts, and Σ and S ∗ × S-indexed set Σ = {Σw,s }w,s∈S ∗ ×S of operation symbols, where S ∗ denotes the free monoid generated by S. We denote each f ∈ Σw,s as f : w → s. In particular, a constant of sort s is an operation a :  → s, with  the empty word. An order-sorted (OS) signature is a triple Σ = (S, ≤, Σ) with (S, ≤) a poset and (S, Σ) a many-sorted signature. S = S/ ≡≤ , the quotient of S under the equivalence relation ≡≤ = (≤ ∪ ≥)+ , is called the set of connected components of (S, ≤). Note that a many-sorted signature Σ is the special case where the poset (S, ≤) is discrete, i.e., s ≤ s iff s = s . The order ≤ and equivalence ≡≤ are extended to sequences of same length in the usual way, e.g., s1 . . . sn ≤ s1 . . . sn iff si ≤ si , 1 ≤ i ≤ n. Σ is called

Order-Sorted Rewriting and Congruence Closure

497

sensible1 if for any two f : w → s, f : w → s ∈ Σ, with w and w of same length, we have w ≡≤ w ⇒ s ≡≤ s . For connected components [s1 ], . . . , [sn ], [s] ∈ S [s ]...[sn ]

f[s]1

= {f : s1 . . . sn → s ∈ Σ | si ∈ [si ], 1 ≤ i ≤ n, s ∈ [s]}

denotes the family of “subsort polymorphic” operators f .



Definition 2. For Σ = (S, Σ) a many-sorted signature, a Σ-algebra is an S-indexed set A = {As }s∈S together with an assignment of: (i) to each constant a :  → s of sort s an element Aa ∈ As , and (ii) to each operation f : w → s, with w = s1 . . . sn , n ≥ 1, a function Af :w→s : Aw → As , where, by convention, As1 ...sn = As1 × . . . × Asn . For Σ = (S, ≤, Σ) an OS signature, an order-sorted Σ-algebra A is a manysorted (S, Σ)-algebra A such that: – whenever s ≤ s , then we have As ⊆ As , and  [s ]...[sn ] and a ∈ Aw ∩ Aw , then we have – whenever f : w → s, f : w → s ∈ f[s]1 Af :w→s (a) = Af :w →s (a). A many-sorted Σ-homomorphism h : A → B is an S-indexed family of functions h = {hs : As → Bs }s∈S such that: (i) for a :  → s, hs (Aa ) = Ba , and (ii) for f : w → s with w = , Af ; hs = hw ; Bf . An order-sorted Σ-homomorphism h : A → B is a many-sorted (S, Σ)homomorphism such that whenever [s] = [s ] and a ∈ As ∩ As , then we have hs (a) = hs (a). We call h injective, resp. surjective, resp. bijective, iff for each s ∈ S hs is injective, resp. surjective, resp. bijective. We call h an isomorphism if there is another order-sorted Σ-homomorphism g : B → A such that for each s ∈ S, hs ; gs = 1As , and gs ; hs = 1Bs , with 1As , 1Bs the identity functions on As , Bs . This defines a category OSAlgΣ .  Theorem 1 [21]. The category OSAlgΣ has an initial algebra. Furthermore, if Σ is sensible, then the term algebra TΣ with: – if a :  → s then a ∈ TΣ,s ( denotes the empty string), – if t ∈ TΣ,s and s ≤ s then t ∈ TΣ,s , – if f : s1 . . . sn → s and ti ∈ TΣ,si 1 ≤ i ≤ n, then f (t1 , . . . , tn ) ∈ TΣ,s , is initial, i.e., there is a unique Σ-homomorphism to each Σ-algebra.  TΣ,[s] denotes the set TΣ,[s] =   For [s] ∈ S, s ∈[s] TΣ,s . Similarly, TΣ will (ambiguously) denote both the above-defined S-sorted set and the set TΣ = 1

The notion of a sensible signature is a minimal syntactic requirement to avoid excessive ambiguity. For example, a many-sorted signature Σ with sorts A, B and C, constant a :  → A and operations f : A → B and f : A → C is not sensible and therefore is intrinsically ambiguous: the term f (a) has both sorts B and C, which are completely different sorts.

498

J. Meseguer



s∈S TΣ,s . We say that an OS signature Σ has non-empty sorts iff for each s ∈ S, TΣ,s = ∅. We will assume throughout that Σ has non-empty sorts. An S-sorted set X = {Xs }s∈S of variables, satisfies s = s ⇒ Xs ∩ Xs = ∅, and the variables in X are always assumed disjoint from all constants in Σ. The Σ-term algebra on variables X, TΣ (X), is the initial algebra for the signature Σ(X) obtained by adding to Σ the variables X as extra constants. Since a Σ(X)algebra is just a pair (A, α), with A a Σ-algebra, and α an interpretation of the constants in X, i.e., an S-sorted function α ∈ [X→A], the Σ(X)-initiality of TΣ (X) can be expressed as the following corollary of Theorem 1:

Theorem 2 (Freeness Theorem). If Σ is sensible, for each A ∈ OSAlgΣ and α ∈ [X→A], there exists a unique Σ-homomorphim, α : TΣ (X) −→ A extending α, i.e., such that for each s ∈ S and x ∈ Xs we have xαs = αs (x). The first-order language of equational Σ-formulas 2 is defined in the usual way: its atoms are Σ-equations t = t , where t, t ∈ TΣ (X)[s] for some [s] ∈ S and each Xs is assumed countably infinite. The set Form(Σ) of equational Σ-formulas is then inductively built from atoms by: conjunction (∧), disjunction (∨) negation (¬), and universal (∀x:s) and existential (∃x:s) quantification with sorted variables x:s ∈ Xs for some s ∈ S. The literal ¬(t = t ) is denoted t = t . Given a Σ-algebra A, a formula ϕ ∈ Form(Σ), and an assignment α ∈ [Y →A], with Y = fvars(ϕ) the free variables of ϕ, we define the satisfaction relation A, α |= ϕ inductively as usual: for atoms, A, α |= t = t iff tα = t α; for Boolean connectives it is the corresponding Boolean combination of the satisfaction relations for subformulas; and for quantifiers: A, α |= (∀x : s) ϕ (resp. A, α |= (∃x : s) ϕ) holds iff for all a ∈ As (resp. some a ∈ As ) we have A, α  {(x:s, a)} |= ϕ, where the assignment α  {(x:s, a)} extends α by mapping x:s to a. Finally, A |= ϕ holds iff A, α |= ϕ holds for each α ∈ [Y →A], where Y = fvars(ϕ). We say that ϕ is valid (or true) in A iff A |= ϕ. We say that ϕ is satisfiable in A iff ∃α ∈ [Y →A] such that A, α |= ϕ, where Y = fvars(ϕ). An order-sorted equational theory is a pair T = (Σ, E), with E a set of Σ-equations. OSAlg(Σ,E) denotes the full subcategory of OSAlgΣ with objects those A ∈ OSAlgΣ such that A |= E, called the (Σ, E)-algebras. OSAlg(Σ,E) has an initial algebra TΣ/E [21], further discussed in Sect. 3. Given T = (Σ, E) and ϕ ∈ Form(Σ), we call ϕ T -valid, written E |= ϕ, iff A |= ϕ for each A ∈ OSAlg(Σ,E) . We call ϕ T -satisfiable iff there exists A ∈ OSAlg(Σ,E) with ϕ satisfiable in A. Note that ϕ is T -valid iff ¬ϕ is T -unsatisfiable. Σ = ((S, ≤), Σ) is a subsignature of Σ  = ((S  , ≤ ), Σ  ), denoted Σ ⊆ Σ  , iff (S, ≤) ⊆ (S  , ≤ ) is a subposet inclusion, and Σ ⊆ Σ  . A signature map H : Σ → Σ  is a monotonic function H : (S, ≤) → (S  , ≤ ) of the underlying posets of sorts together with a mapping H : Σ  (f : s1 . . . sn → s) → (H(f ) : 2

There is only an apparent lack of predicate symbols. To express a predicate p(x1 : s1 , . . . , xn:sn ), add a new sort Truth with a constant tt, and with {Truth} a separate connected component, and view p as a function symbol p : s1 , . . . , sn → Truth. An atomic formula p(t1 , . . . , tn ) is then expressed as the equation p(t1 , . . . , tn ) = tt.

Order-Sorted Rewriting and Congruence Closure

499

H(s1 ) . . . H(sn ) → H(s)) ∈ Σ  . H induces a map H : Form(Σ) → Form(Σ  ). A signature inclusion Σ ⊆ Σ  defines a signature map Σ → Σ  : f → f . A signature map H : Σ → Σ  induces a functor in the opposite direction |H : OSAlgΣ   B → B |H ∈ OSAlgΣ , where the H-reduct B |H has: (i) for each s ∈ S, (B |H )s = BH(s) ; and (ii) for each f : s1 . . . sn → s in Σ, (B |H )f = BH(f ) . For H : Σ → Σ  a signature inclusion, B |H is denoted B |Σ . For B ∈ OSAlgΣ  and ϕ ∈ Form(Σ) with fvars(ϕ) = ∅ we have [21]: (†)

3

B |= H(ϕ) ⇔ B |H |= ϕ.

Order-Sorted Rewriting and Equality

Given an OS signature Σ = ((S, ≤), Σ), a Σ-rewrite rule 3 is a sequent l → r  An order-sorted term rewriting system with l, r ∈ TΣ (X)[s] for some [s] ∈ S. (OSTRS) is then a pair (Σ, R) with R a set of Σ-rewrite rules. Since, as shown in the Introduction, replacement of equals for equals and standard rewriting break down in the order-sorted case, we should define rewriting deductions with an OSTRS not by means of the reflexive-transitive closure →∗R of the rewrite relation →R , but by means of an inference system with two  corresponding kinds of sequents: sequents t → t , where t, t ∈ TΣ (X)[s] , [s] ∈ S,   to one-step application of rules, and sequents t → t , where t, t ∈ TΣ (X)[s] ,  corresponding to more complex rewriting deductions. The symbol → [s] ∈ S, is close enough to →∗ to suggest that: (i) it plays a role similar to a reflexive transitive-closure in the unsorted case, but (ii) in general it is different from such a closure. For example, for Σ the signature in the Introduction and R = {a → b, b → c}, we can derive f (a) → f (c), but there is no sequence of one-step rewrites from f (a) to f (c). We then define two kinds of rewriting deductions: (Σ, R) t → t and (Σ, R) t → t , as those sequents derivable from (Σ, R) by a finite application of the following inference rules, where σ denotes an S-sorted substitution, i.e., an S-sorted function σ ∈ [X→TΣ (X)]: Reflexivity Subsumption Transitivity Congruence

Replacement 3

t → t t → t t → t t → t t → t   t→ t u1 → u1 . . . un → un f (u1 , . . . , un ) → f (u1 , . . . , un ) where f (u1 , . . . , un ), f (u1 , . . . , un ) ∈ TΣ (X) tσ → t σ where t → t ∈ R

For greater generality no restriction is placed on the variables of l and r.

500

J. Meseguer

The first three and the last inference rule are standard, but the Congruence rule is more subtle. We can better understand these rules by means of our running example (Σ, R). The sequent f (a) → f (b) is not derivable: the attempt to obtain it by applying Replacement with rule a → b, Subsumption to get a → b, and then Congruence fails, because of the side condition, since f (b) ∈ TΣ (X). To see what can be derived, consider the derivation of the sequent f (a) → f (c). Since we have rules a → b and b → c, we can derive a → c by two applications of Replacement followed by Subsumption and one application of Transitivity. Then Congruence gives us: a → c f (a) → f (c) Note the interesting fact that f (a) is typed with f : A → A, and f (c) is typed with f : C → C. We can think of Congruence as a “tunneling rule.” f (a) → f (c) cannot be obtained by composing one-step rewrites: failed attempts such as that for deriving f (a) → f (b) make it impossible; but we can “tunnel through” such failed attempts and obtain a more complex sequent like f (a) → f (c) when the left- and right-hand sides are well-formed terms in TΣ (X). The above inference system yields as a special case a sound and complete inference system for order-sorted equational logic: we just view an order-sorted equational theory (Σ, E) as the OSTRS (Σ, R(E)), where R(E) = {t → t | t = t ∈ E ∨ t = t ∈ E}. That is, equality steps are viewed as either left-to-right or right-to-left rewrite steps. We then have: Definition 3. Given an order-sorted equational theory (Σ, E) with Σ sensible, its equational deduction relation, denoted (Σ, E) u = v, or just E u = v, is defined by the equivalence: (Σ, E) u = v



(Σ, R(E)) u → v.

Theorem 3 (Soundness and Completeness) [21] Theorem 24. For Σ sensible and E ∪ {u = v} a set of Σ-equations we have the equivalence: (Σ, E) u = v



(Σ, E) |= u = v

The above theorem has as a corollary the construction of the initial algebra TΣ/E for the category OSAlg(Σ,E) of (Σ, E)-algebras. Assuming Σ sensible, TΣ/E , has an easy definition. Note that the relation E u = v induces an  We then define TΣ/E,s = equivalence relation =E on each set TΣ,[s] , [s] ∈ S.  {[t]=E ∈ TΣ,[s] / =E | [t]=E ∩TΣ,s = ∅} for each s ∈ [s], and define each operation f : s1 . . . sn → s ∈ Σ by the map ([t1 ]=E , . . . , [t1 ]=E ) → [f (t1 , . . . , tn )]=E , where ti ∈ [ti ]=E ∩ TΣ,si , 1 ≤ i ≤ n, showing it does not depend on the choice of ti ’s. 3.1

Kind-Complete OS-Rewriting and Equational Deduction

The order-sorted rewrite relation t → t is obviously quite impractical and hard to implement. For this reason, given an OSTRS (Σ, R) several conditions on

Order-Sorted Rewriting and Congruence Closure

501

either Σ or R have been sought to be able to perform rewriting computations in essentially the standard and efficient way in which it is performed in an unsorted or many-sorted TRS. Two such conditions, going back to [14], are to either: (i) require that the rules R are sort-decreasing, i.e., for each l → r ∈ R and S-sorted substitution σ, if lσ ∈ TΣ,s then rσ ∈ TΣ,s (this can be checked by the method explained in [17]); or (ii) if R is not sort-decreasing, extend Σ with new “retract operators” rs,s : s → s , s, s ∈ [s], s ≤ s , to catch typing errors, add to R “error recovery” rules of the form rs,s (x:s ) → x:s , and force sort-decreasingness of R by replacing each not sort-decreasing u → v ∈ R by suitable rules of the form uσ → rs,s (vσ), where σ may lower the sorts of some variables. Conditions (i) or in its defect (ii) work and can be shown to be conservative in a certain sense [14]. However, they have serious limitations. Sort decreasingness is a strong condition that may be impossible to achieve for some OSTRS arising in practice; and if the solution with retracts is adopted, an unpleasant consequence is that we change the models, including the initial ones, since retracts add new operations and new error terms to the original sorts. All these limitations can be avoided —while allowing rewriting with rules R and equational deduction with equations E to be performed in the standard way— by using a faithful embedding of order-sorted equational logic into membership equational logic (MEL) [3, 21]. MEL introduces a typing distinction between sorts s ∈ S, which may be related by subsort relations just as in the order-sorted way, and the kind [s] associated to each connected component  which is above all sorts in [s]. An ill-formed term like f (b) in the OS [s] ∈ S, signature of the Introduction has no sort, but has kind [B] . In this way, the earlier side condition in the Congruence rule in Sect. 3 can be avoided. The faithfulness of this embedding of logics means in particular that both initial models and equational deduction are preserved ([21], Corollary 28). However: (i) the proof in [21] is model-theoretic; (ii) it focuses on the equational logic level, and does not deal with the more general rewriting logic level; and (iii) it assumes that the entire MEL framework is adopted. Can the essential advantages of this embedding be still obtained while remaining at the ordersorted level ? The answer is yes! Since: (i) this solution plays a key role in the treatment of satisfiability for the theory of OS uninterpreted function symbols in Sect. 4, and (ii) having a much simpler theory of OS rewriting is useful in its own right, I give a detailed treatment of it below. The key idea is to use a signature transformation Σ → Σ  extending any OS signature Σ into one whose components have a top sort, understood as the kind of that component. The essential point is that Σ  belongs to a class of order-sorted signatures called kind complete where both rewriting and equational deduction can be performed in the standard way. Definition 4. An OS signature Σ = ((S, ≤), Σ) is called kind-complete iff each connected component [s] ∈ S has a top sort [s] , called its kind, with [s] ≥ s [s ]...[sn ] for each s ∈ [s], and any non-empty subsort-polymorphic family f[s]1 ⊆Σ

502

J. Meseguer

includes the typing f : [s1 ] , . . . , [sn ] → [s] . Note that any many-sorted Σ — and in particular any unsorted (i.e., single-sorted) Σ— is trivially kind-complete. Any OS signature Σ can be extended to a kind-complete one by a transformation Σ → Σ  . Σ  is constructed in two-steps: (i) we first associate to the  = (S , Σ),  where order-sorted signature ((S, ≤), Σ) the many-sorted signature Σ [s ]...[s 1 n]  and with f : [s ] . . . [s ] → [s] ∈ Σ  iff f = ∅; S = {[s] | [s] ∈ S}, 1 n [s]  2   and (ii) we then define Σ = ((S  S , ≤ ), Σ  Σ), where ≤ ∩S = ≤, and for each [s] ∈ S we have s } } >> eˆ0 } > e+ g fˆ >>  ~ }}}  ~} pp Aˆ o o qˆ

ϕV ˆ X

/ / Tˆ X ˆ V h+ V

 ˆ B

This proves the theorem for the unordered case. The proof for the ordered case is analogous: replace profinite equations by inequations, and use the homomorphism theorem for ordered algebras to construct the morphism g.

Profinite Monads, Profinite Equations, and Reiterman’s Theorem

5

543

Applications and Examples

Let us consider some examples and applications. First note that the original Reiterman theorem and its ordered version emerge from Theorem 4.12 by taking the identity monad T = Id and X = free finitely generated objects of D, see  T  and a profinite  = Id, D T = D, D = D, Example 4.8(a). In this case we have T equation u = v (resp. a profinite inequation u ≤ v) is a pair u, v ∈ ΦˆX for a finite set X. We conclude: Corollary 5.1 (Reiterman [23], Banaschewski [6]). Let D be a variety of algebras. A class V ⊆ Df is a pseudovariety iff it is presented by profinite equations over finite sets of variables. Corollary 5.2 (Pin and Weil [20]). Let D be a variety of ordered algebras. A class V ⊆ Df is a pseudovariety iff it is presented by profinite inequations over finite sets of variables. Recall from Isbell [13] that a class V ⊆ D is closed under products and subalgebras iff it is presented by implications  si = ti ⇒ s = t i∈I

where si , ti , s, t are terms and I is a set. Choosing T to be the identity monad and X = Df gives us the counterpart for finite algebras: by Example 4.8(b) a Df -pseudovariety is precisely a class V ⊆ Df closed under finite products and subalgebras, since the closure under split quotients is implied by closure under subalgebras. Such a class could be called “quasi-pseudovariety”, but to avoid this clumsy terminology we prefer “quasivariety of finite algebras”. Definition 5.3. A quasivariety of finite algebras of D is a class V ⊆ Df closed under finite products and subalgebras. In analogy to Isbell’s result we show that quasivarieties of finite algebras are precisely the classes of finite algebras of D presented by profinite implications. Definition 5.4. Let X be a finite set of variables. 1. Unordered case: a profinite implication over X is an expression  u i = vi ⇒ u = v

(5.1)

i∈I

where I is a set and ui , vi , u, v ∈ ΦˆX . An object A ∈ Df satisfies (5.1) if for ˆ i ) = h(v ˆ i ) for all i ∈ I one has h(u) ˆ ˆ every h : ΦX → A with h(u = h(v). 2. Ordered case: a profinite implication over X is an expression  u i ≤ vi ⇒ u ≤ v (5.2) i∈I

where I is a set and ui , vi , u, v ∈ ΦˆX . An object A ∈ Df satisfies (5.2) if for ˆ i ) = h(v ˆ i ) for all i ∈ I one has h(u) ˆ ˆ every h : ΦX → A with h(u ≤ h(v).

544

L.-T. Chen et al.

A class P of profinite implications presents the class of all finite algebras in D satisfying all implications in P . Theorem 5.5. For any class V ⊆ Df the following statements are equivalent: 1. V is a quasivariety of finite algebras. 2. V is presented by profinite (in-)equations over Df . 3. V is presented by profinite implications. Proof Sketch. 3⇒1 requires a routine verification, and 1⇒2 is Theorem 4.12. For 2⇒3 assume w.l.o.g. that V is presented by a single profinite equation u = v with u, v elements of some X ∈ Df . Express X as a quotient q : ΦY  X for some finite set Y . Let { (ui , vi ) : i ∈ I } be the kernel of qˆ: ΦˆY  X (consisting of all pairs (ui , vi ) ∈ ΦˆY × ΦˆY with qˆ(ui ) = qˆ(vi )), and choose u , v  ∈ ΦˆY with qˆ(u ) = u and qˆ(v  ) = v. Then a finite object A ∈ Df satisfies the profinite equation u = v iff it satisfies the profinite implication  u i = vi ⇒ u  = v  , (5.3) i∈I

which proves that V is presented (5.3). Analogously for the ordered case. Example 5.6. 1. Let V ⊆ Monf be the quasivariety of all finite monoids whose only invertible element is the unit. It is presented by the profinite implication xω = 1 ⇒ x = 1 over the set of variables X = {x}. Here the profinite ∗ is interpreted, for every finite monoid M with x interpreted word xω ∈ X as m ∈ M , as the unique idempotent power of m. Indeed, if M has no nontrivial invertible elements, it satisfies the implication: given m = 1 and mk idempotent, then mk = 1 (otherwise m has the inverse mk−1 ). Conversely, if M satisfies the implication and m is invertible, then so is its idempotent power mk . Hence mk · mk = mk implies mk = 1, so m = 1. 2. Let Pos be the variety of posets (i.e. the variety of all ordered algebras over the empty signature). The quasivariety V ⊆ Posf of finite discrete posets is presented by the profinite implication v ≤ u ⇒ u ≤ v over the set X = {u, v}. Remark 5.7. As indicated before all concepts in this paper also apply to a setting where D is a many-sorted variety of algebras or ordered algebras. In this case an algebra is finite if the disjoint union of the underlying sets of all sorts is a finite set. By a profinite equation over X ∈ D is a meant pair of elements ˆ and it is satisfied by a finite T-algebra A if for every u, v in some sort s of TˆX, ˆ → A merges u, v. T-homomorphism h : T X → A the s-component of h+ : TˆX Similarly for profinite inequations and profinite implications. Example 5.8. Consider the variety D of directed graphs, i.e. algebras for the two-sorted signature consisting of a sort Ob (objects), a sort Mor (morphisms) and two unary operations s, t : Mor → Ob specifying the source and target of a morphism. Then Cat, the category of small categories and functors, is isomorphic to D T for the monad T constructing the free category on a graph. Choosing

Profinite Monads, Profinite Equations, and Reiterman’s Theorem

545

X = free finitely generated graphs, Theorem 4.12 shows that every pseudovariety of categories, i.e. every class of finite categories closed under finite products, subcategories (represented by injective functors) and quotient categories (represented by surjective functors), can be specified by profinite equations over a two-sorted set of variables. This result was essentially proved by Jones [15]. The difference is that he restricts to quotients represented by surjective functors which are bijective on objects, and replaces subcategories by faithful functors. Moreover, profinite equations are restricted to the sort of morphisms.

6

Conclusions and Future Work

Motivated by recent developments in algebraic language theory, we generalised Reiterman’s theorem to finite algebras for an arbitrary monad T on a base category D. Here D is a variety of (possibly ordered, many-sorted) algebras.  of T, which makes it The core concept of our paper is the profinite monad T possible to introduce profinite (in-)equations at the level of monads and prove that they precisely present pseudovarieties of T-algebras. Referring to the diagram in the Introduction, our Reiterman theorem is presented in a setting that unifies the two categorical approaches to algebraic language theory of Boja´ nzcyk [9] and in our work [1–3, 10]. The next step is to also derive an Eilenberg theorem in this setting. For each monad T on a category of sorted sets, Boja´ nczyk [9] proved an Eilenberg-type characterisation of pseudovarieties of T-algebras: they correspond to varieties of T-recognisable languages. Here by a “language” is meant a function from T X to {0, 1} for some alphabet X, and a variety of languages is a class of such languages closed under boolean operations, homomorphic preimages and a suitably generalised notion of derivatives. On the other hand, as indicated in Example 2.12, one needs to consider monoids on algebraic categories beyond Set in order to study varieties of languages with relaxed closure properties, e.g. dropping closure under complement or intersection. The aim is thus to prove an Eilenberg theorem parametric in a monad T on an algebraic category D. Observing that e.g. for D = Set  on Stone dualises to a comonad on the category of boolean algethe monad T bras, we expect this can be achieved in a duality-based setting along the lines of Gehrke, Grigorieff and Pin [12] and our work [1, 3]. Throughout this paper we presented the case of ordered and unordered algebras as separated but analogous developments. Pin and Weil [20] gave a uniform treatment of ordered and unordered algebras by generalising Reiterman’s theorem from finite algebras to finite first-order structures. A similar approach should also work in our categorical framework: replace D by a variety of relational algebras over a quasivariety Q of relational first-order structures, with Q = Set and Q = Pos covering the case of algebras and ordered algebras. Finally, observe that categories of the form D T , where D is a many-sorted variety of algebras and T is an accessible monad, correspond precisely to locally presentable categories. This opens the door towards an abstract treatment, and further generalisation, of Reiterman’s theorem in purely categorical terms.

546

L.-T. Chen et al.

References 1. Ad´ amek, J., Milius, S., Myers, R.S.R., Urbat, H.: Generalized Eilenberg theorem I: local varieties of languages. In: Muscholl, A. (ed.) FOSSACS 2014 (ETAPS). LNCS, vol. 8412, pp. 366–380. Springer, Heidelberg (2014) 2. Ad´ amek, J., Milius, S., Urbat, H.: Syntactic monoids in a category. In: Moss, L.S., Sobocinski, P. (eds.) Proceedings of CALCO 2015 (2015) 3. Ad´ amek, J., Myers, R.S.R., Urbat, H., Milius, S.: Varieties of languages in a category. In: Proceedings LICS 2015. IEEE (2015) 4. Ad´ amek, J., Rosick´ y, J.: Locally Presentable and Accessible Categories. Cambridge University Press, Cambridge (1994) 5. Almeida, J.: On pseudovarieties, varieties of languages, filters of congruences, pseudoidentities and related topics. Algebra Universalis 27(3), 333–350 (1990) 6. Banaschewski, B.: The Birkhoff theorem for varieties of finite algebras. Algebra Universalis 17(1), 360–368 (1983) 7. Barr, M.: HSP subcategories of Eilenberg-Moore algebras. Theory Appl. Categ. 10(18), 461–468 (2002) 8. Bloom, S.L.: Varieties of ordered algebras. J. Comput. Syst. Sci. 13(2), 200–212 (1976) 9. Boja´ nczyk, M.: Recognisable languages over monads. In: Potapov, I. (ed.) DLT 2015. LNCS, vol. 9168, pp. 1–13. Springer, Heidelberg (2015). Full version: http://arxiv.org/abs/1502.04898 10. Chen, L.T., Urbat, H.: A fibrational approach to automata theory. In: Moss, L.S., Sobocinski, P. (eds.) Proceedings of CALCO 2015 (2015) 11. Eilenberg, S.: Automata, Languages, and Machines, vol. 2. Academic Press, New York (1976) ´ Duality and equational theory of regu12. Gehrke, M., Grigorieff, S., Pin, J.E.: lar languages. In: Aceto, L., Damg˚ ard, I., Goldberg, L.A., Halld´ orsson, M.M., Ing´ olfsd´ ottir, A., Walukiewicz, I. (eds.) ICALP 2008, Part II. LNCS, vol. 5126, pp. 246–257. Springer, Heidelberg (2008) 13. Isbell, J.R.: Subobjects, Adequacy, Completeness and Categories of Algebras. Instytut Matematyczny Polskiej Akademi Nauk, Warsaw (1964) 14. Johnstone, P.T.: Stone Spaces. Cambridge University Press, Cambridge (1982) 15. Jones, P.R.: Profinite categories, implicit operations and pseudovarieties of categories. J. Pure Appl. Algebr. 109(1), 61–95 (1996) 16. Kennison, J.F., Gildenhuys, D.: Equational completion, model induced triples and pro-objects. J. Pure Appl. Algebr. 1(4), 317–346 (1971) 17. Linton, F.E.J.: An outline of functorial semantics. In: Eckmann, B. (ed.) Seminor on Triples and Categorical Homology Theory. LNM, vol. 80, pp. 7–52. Springer, Heidelberg (1969) 18. Lane, S.M.: Categories for the Working Mathematician, 2nd edn. Springer, New York (1998) 19. Pin, J.E.: A variety theorem without complementation. Russ. Math. (Izvestija vuzov.Matematika) 39, 80–90 (1995) 20. Pin, J.E., Weil, P.: A Reiterman theorem for pseudovarieties of finite first-order structures. Algebra Universalis 35, 577–595 (1996) 21. Pol´ ak, L.: Syntactic semiring of a language. In: Sgall, J., Pultr, A., Kolman, P. (eds.) MFCS 2001. LNCS, vol. 2136, p. 611. Springer, Heidelberg (2001) 22. Priestley, H.A.: Ordered topological spaces and the representation of distributive lattices. Proc. London Math. Soc. 3(3), 507 (1972)

Profinite Monads, Profinite Equations, and Reiterman’s Theorem

547

23. Reiterman, J.: The Birkhoff theorem for finite algebras. Algebra Universalis 14(1), 1–10 (1982) 24. Reutenauer, C.: S´eries formelles et alg`ebres syntactiques. J. Algebra 66, 448–483 (1980) 25. Ribes, L., Zalesskii, P.: Profinite Groups. A Series of Modern Surveys in Mathematics. Springer, Berlin (2010) 26. Sch¨ utzenberger, M.P.: On finite monoids having only trivial subgroups. Inf. Control 8, 190–194 (1965) 27. Wilke, T.: An Eilenberg theorem for infinity-languages. In: Proceedings of ICALP 1991, pp. 588–599 (1991)

Author Index

Abdulla, Parosh Aziz 319 Adámek, Jiří 531 Ahman, Danel 36 Amtoft, Torben 180 Aubert, Clément 373 Axelsen, Holger Bock 73

Kaarsgaard, Robin 73 Katsumata, Shin-ya 513 Kesner, Delia 424 Kiefer, Stefan 303 Krčál, Jan 353 Krishna, Shankara Narayanan

Bagnol, Marc 373 Banerjee, Anindya 180 Birkedal, Lars 20 Bizjak, Aleš 20 Brenguier, Romain 217

Laird, James 55 Lasota, Sławomir 445 Lazić, Ranko 162, 445 Leroux, Jérôme 445

Capretta, Venanzio 91 Chaudhuri, Kaustuv 390 Chen, Liang-Ting 531 Chistikov, Dmitry 252, 462 Ciobanu, Radu 319 Clouston, Ranald 20 Czerwiński, Wojciech 462 Decker, Normann 269 Dräger, Klaus 479 Durand-Gasselin, Antoine Fijalkow, Nathanaël 303 Fortin, Marie 197 Francalanza, Adrian 145 Fujii, Soichiro 513 Gastin, Paul 197 Ghani, Neil 3, 36 Ghilardi, Silvio 126 Gouveia, Maria João 126 Grathwohl, Hans Bugge 20 Habermehl, Peter 285 Hermanns, Holger 353 Hofman, Piotr 445, 462

Madnani, Khushraj 335 Marin, Sonia 390 Martyugin, Pavel 252 Mayr, Richard 319 Melliès, Paul-André 513 Meseguer, José 493 Milius, Stefan 107, 531 Møgelberg, Rasmus E. 20 Murawski, Andrzej S. 162 Nordvall Forsberg, Fredrik

285

335

3

Pagani, Michele 408 Pandya, Paritosh K. 335 Pattinson, Dirk 107 Pierron, Théo 234 Pilipczuk, Michał 462 Place, Thomas 234 Plotkin, Gordon D. 36 Sangnier, Arnaud 319 Santocanale, Luigi 126 Schmitz, Sylvain 445 Seiller, Thomas 373 Shirmohammadi, Mahsa 252, 303 Simpson, Alex 3 Sproston, Jeremy 319 Straßburger, Lutz 390

550

Author Index

Tasson, Christine 408 Thoma, Daniel 269 Totzke, Patrick 445

Vaux, Lionel 408 Vester, Steen 353 Wehar, Michael 462 Wißmann, Thorsten 107

Urbat, Henning 531 Uustalu, Tarmo 91

Zeitoun, Marc

234